Machine learning for Marketers

Machine Learning
for Marketers
A COMPREHENSIVE GUIDE TO MACHINE LEARNING

Introduction
CH 1
The Basics of Machine Learning
CH. 2
Supervised vs Unsupervised Learning and
Other Essential Jargon
CH. 3
What Marketers can Accomplish
with Machine Learning
CH. 4
Successful Machine Learning Use Cases
CH. 5
How Machine Learning Guides SEO
CH. 6
Chatbots: The Machine Learning you
are Already Interacting with
CH. 7
How to Set Up a Chatbot
CH. 8
How Marketers Can Get Started with Machine
Learning
CH. 9
Most Effective Machine Learning Models
CH. 10
How to Deploy Models Online
CH. 11
How Data Scientists Take Modeling
to the Next Level
CH. 12
Common Problems with Machine Learning
CH. 13
Machine Learning Quick Start
pg 3
pg 4
pg 9
pg 13
pg 18
pg 26
pg 30
pg 36
pg 45
pg 58
pg 65
pg 72
pg 79
pg 84
CONTENTS

3
Machine learning is a term thrown around in technol-
ogy circles with an ever-increasing intensity. Major
technology companies have attached themselves to
this buzzword to receive capital investments, and every
major technology company is pushing its even shinier
parentartificial intelligence (AI).
The reality is that Machine Learning as a concept is as
old as computing itself. As early as 1950, Alan Turing was
asking the question, “Can computers think?” In 1969,
Arthur Samuel helped define machine learning specifi-
cally by stating, “[Machine Learning is the] field of study
that gives computers the ability to learn without being
explicitly programmed.”
While the concept has been around for more than half a
century, we have finally reached a point in technological
advancement where hardware and software can actually
help developers match their aspirations with tangible
reality. This development has led to not only the rise
of machine learning and AI advancements, but, more
importantly, also advancements inexpensive enough for
anyone to use.
Why Did We Create This Guide?
While the topic of machine learning and AI has been
exhaustively covered in the technology space, a singu-
lar comprehensive guide has not been created in the
marketing space on the topic, including how it affects
marketers and their work. This space is so thick with
technology-based terminology, but not every marketer
has the chops to venture into the space with confi-
dence. With many products coming to market, iPullRank
believes that preparing marketers to tackle the land-
scape armed with a solid foundation is important.
Who Is This Guide For?
Since machine learning touches an ever-increasing
number of industries, we’ll also touch on several differ-
ent ways that machine learning is impacting people in
many professions. Most data scientists use R and Python
for machine learning, but have you met a marketer these
days that only lives and breathes data science? We cre-
ated this guide for the marketers among us whom we
know and love by giving them simpler tools that don’t
require coding for machine learning.
We’ll briefly touch on other spaces to round out market-
ers’ understanding of the complex topic we’re tackling.
We’ll also look at how machine learning can help mar-
keters through examination of use cases, plus we’ll
dive into martech, a field that’s increasingly including
machine-learning concepts.
How Will This Guide Help?
Since we want to create the most comprehensive
resource on machine learning for marketers, this guide
will be far more than explanatory. We’ll be looking at rel-
evant use cases related to machine learning and delving
into practical use of machine learning so that you can
begin to use the technologies we’ll discuss after you read
this guide.
We’ll also discuss essential jargon that you’ll need to know
about machine learning and how machine learning guides
SEO specifically, plus we’ll delve into the topic of chatbots
(just what do they do, anyway?).
Finally, we’ll help marketers actually use machine learning,
but focus on the common problems beginners face. Along
the way, you’ll get a look at specific tools and platforms
you can use more effectively.
Machine learning is a vital tool for marketers to add to
their knowledge base and future-proof their skill sets. As
the technology behind it continues to develop, machine
learning won’t be something you read about in tech arti-
cles; it’ll be essential to organizations of all sizes.
INTRODUCTION

The Basics of
Machine Learning
CHAPTER 1

5
By 2020, the digital universe will be 40,000 exabytes, or 40tn gigabytes, in
comprehensive size. In contrast, the human brain can hold only 1 million
gigabytes of memory. Too much data exists for humans to parse, analyze,
and understand. Here is where machine learning is finding its value: The
raw amount and constant growth of data creates a need for methods to
make sense of that data overload in ways that can impact an array of pro-
fessions and lifestyles.
Although it has many uses, machine learning usually gets deployed to solve problems by finding pat-
terns in data we can’t see ourselves. Computers give us the power to unearth concepts that are either
too complex for humans or would take us longer to than we’d like to practically use them as a solution.
The Basic Machine Learning Model
The first step in machine learning is identifying the rules. The automation, or machine part, comes
secondary. Rules are essentially the logic upon which we build the automation.
The first step in rule creation is finding the basic breakdown of the data you want to learn about. Think
of this area as the labels you give your data in an Excel sheet or database.
Google called these labels “Parameters: the signals or factors used by the model to form its
decisions” during a Machine Learning 101 event it held in 2015. A good example here would be work-
ing with stock prices to see how different variables can affect the market. In our case, the Parameters
would be the stock price, the dates, and the company.
Next, identify the positive and negative results your automation looks to unearth. Essentially, think of
this idea in terms of programmatic words such as “True” and “False.” You need to essentially “teach”
your program how you, and it, should evaluate your data. Google calls this teaching the “Model: the
system that makes predictions or identifications.”
Based on our example we used for the “Parameters,” let’s look at how a basic setup of the “model”
would look.
• If we want to see how different variables would affect stock prices, a human being
would need to assign the logic of dates and prices with the variables that affected
them, such as the upticks in stocks post-war and in conflicts.
• Once you create the basic logic, you take that logic and your data parameters begin to
grow your data set you intend to use in the learning stage. At this point in the process
you may find your data wanting, and this is the reason to not begin the process data first.
• From here, you run the data through algorithms and tools to solve the logic created.
Google calls this process the “Learner: the system that adjusts the parameters —
and in turn the model — by looking at differences in predictions versus actual
outcome.” In our stock example, our learner would be looking for what variables could
have possible impacts on the stocks we’re looking to buy and give us predictions about
whether the data suggests that the purchase is a short- or long-term buy.
• Gradient learning, or gradient decline, are an important part of the process from here.
We’re talking about small adjustments — not large leaps — that the computer makes
over time until it gets the results correct. But you have to watch out for anomalies in
the data; they can have huge impacts on the results.
For marketers, we can find a clear use case we can talk about for this basic description of machine
THE BASICS OF MACHINE LEARNINGCH. 1MACHINE LEARNING FOR MARKETERS

Rules are
essentially
the logic
upon which
we build the
automation.
THE BASICS OF MACHINE LEARNING

7
learning: Google. What would our life be like without it?
Even in its early form, Google took indexed webpages and unstructured data points around them and
arranged them based on logic created using the original PageRank. All of this arrangement happened
because of tools used to create a result: a search engine results page (SERP).
Marketers have been dealing with machine learning in one form or another for some time now.
Facebook feeds, Twitter trends, and algorithmic ad-buying platforms all use a form of machine learn-
ing to make decisions easier.
Supervised Versus Unsupervised Machine Learning
One of the largest fallacies with machine learning is that it’ll replace the need for humans. But didn’t
we just show you above how humans work among several levels of the process? This idea goes beyond
basic data scientists and engineers, extending to people who can shape the problems that machine
learning will solve, extract the results from the learning process, and apply those results in a meaning-
ful way. As we’ve seen with Google quality raters, humans must also qualify the results and help refine
the logic used for learning.
Machine learning is actually a method for human capital enrichment. It can super-charge the results
achieved by marketers and expand the scope of what we even consider positive results and returns.
We can further break apart machine learning into two parts: supervised and unsupervised learning.
• With supervised learning, you deal with labeled data, and you try to optimize your
machine learning algorithm to produce the single, correct output for each input.
• With unsupervised learning, you deal with unlabeled data, so you don’t know what the
output will be.
Instead of training your algorithm to return a correct output based on a specific input, you’re training
it to look for structure and patterns in the entire data set.
Different Variants of Machine Learning
Are you still with us so far? Good, because now we’re getting to some cool stuff.
Artificial Intelligence, Deep Learning, and Natural Language Processing: They’re shock-and-awe words,
but what do they mean? Well, they’re related concepts that have perhaps larger implications on human
capital replacement and interaction.
• Artificial Intelligence (AI): We can look at this concept as a computer doing
something that humans need intelligence to do. Historically, the Turing test was used
to determine whether a computer has a level of intelligence equal to that of a human.
The Turing test invites a participant to exchange messages, in real time, with an unseen
party. In some cases, that unseen party is another human, but in other cases, it’s a
computer. If the participant is unable to distinguish the computer from the human, we
say that the computer has passed the Turing test and is considered intelligent.
However, deeper concepts guide AI development today than merely parroting of human language via
computers. AI research looks to develop technology that takes action based on learned patterns. You
can break down AI into the smaller segments of deep learning — sometimes called neural networks
— and natural language processing.
• Deep learning uses the human brain as its design model. It layers brain-like neurons
created from levels of machine learning. Each level does its learning and produces

8
results that get passed onto the next network that takes on another task with that data.
This process replicates itself so that the program can look at one set of data from an
unlimited number of viewpoints and create an equally unlimited number of solutions.
Google DeepMind’s AlphaGo program — it made headlines when it beat the world’s top-ranked play-
ers at the ancient board game Go — is an example of deep learning in action. The complexities of the
game Go means that a computer can’t simply brute force patterns as you can with a game like chess.
The computer must learn from patterns and use intuition to make choices. This level of operation
isn’t something basic machine learning around a base data set can do; however, deep learning allows
for the layered neurons to rework the data in unlimited ways by looking at every possible solution.
Deep learning is mostly unsupervised and aims to avoid the need for human intervention. At its core,
it looks to learn by watching. When you think about the AlphaGo use case, it all starts to make sense.
• Natural Language Processing (NLP): Here’s where computers use machine learn-
ing to communicate in human language. Search engines and grammar checkers are
both examples of NLP. This type of technology is what people most often associate
with the Turing test. A quality NLP looks to pass the basics of the Turing test, even
though some developers would argue about whether their applications really aim to
“trick” users into thinking their applications are humans. Think about it this way: No
one believes Amazon’s Alexa is a real person answering questions despite its use of
NLP at its core.
Put It All Together
In its current and expanding variation, machine learning is something that is often seen as a confusing
topic. However, as we’ve just described, it breaks neatly into some basic concepts.
• Machine Learning has a clean model for collecting Parameters of data that it feeds
into a human-created Model, which the machine Learner then uses to create or
find a solution. The basic example we used was the original Google PageRank model for
creating SERPs.
• Machine Learning can be unsupervised, often associated with the fields of AI, or
supervised, where humans must create the Models and test quality for the findings of
the Learners.
• Further, advanced Machine Learning, — often identified as AI — breaks into a few
subfields of its own, the most notable of these are Deep Learning and NLP. Deep
Learning uses layered machine learning in an unsupervised way to approach data from
different angles and learn as it moves from layer to layer. NLP uses machine learning to
communicate in languages humans use every day.
We’ll take a deeper look into several of these topics as we move into decoding the industry jargon
associated with machine learning and its variants.

Supervised vs
Unsupervised
Learning and Other
Essential Jargon
CHAPTER 2

10
Diving deeper into the topics surrounding machine learning, we’re con-
fronted with a copious amount of jargon. It helps our journey to understand
how professionals in the space discuss the topics so that we can become
familiarwiththetermswe’llrunintoaswedivedeeperintomachinelearning.
Our goal will be providing an understanding of the various topics without getting too deep into the
technical details. We know your time is important, so we’re making sure that the time spent learning
the language of machine learning will pay off as we go down the path of utilization and use cases.
Revisiting Supervised and Unsupervised Machine Learning
We grazed past the concept of supervised and unsupervised learning in Chapter 1; however, these
topics are important, and they deserve a more in-depth study.
Supervised Machine Learning
As previously discussed, supervised machine learning involves human interaction elements to manage
the machine learning process. Supervised machine learning makes up most of the machine learning
in use.
The easiest way to understand supervised machine learning is to think of it involving an input vari-
able (x) and an output variable (y). You use an algorithm to learn a mapping function that connects
the input to the output. In this scenario, humans are providing the input, the desired output, and
the algorithm.
Let’s look at supervised learning in terms of two types of problems:
Classification – Classification problems use categories as an output variable. Example categories
would be demographic data such as sex or marital status. A common model for these types of prob-
lems are support vector machines. Despite their odd name, support vector machines are a way we
describe a linear decision boundary between classes that maximizes the width of the boundary itself.
Regression – Regression problems are where the output variables are a real number. A common
format of these types of problems are linear progressions. Linear regression models determine the
impact of a number of independent variables on a dependent variable (such as sales) by seeking a
“best fit” that minimizes squared error. Other regression models combine linear models for more
complex scenarios.
One of two basic types of machine learning models uses labeled data to build models that make pre-
dictions. “Labeled” means that the data represents something in the past for which the outcome is
known (for example, an item purchased).
Ensemble Methods – These learning algorithms construct a set of classifiers and then classify new
data points by taking a weighted vote of their predictions (the original ensemble method was known
as the Bayesan average). This mathematical concept estimates the mean of a population using out-
side information, including pre-existing beliefs, much like what you’d do to find a weighted average.
Ensemble methods reduce variance in individual models by combining a number of them together
and averaging predictions.
Logistic Regression – This regression model comes into play when the dependent variable is cate-
gorical. Given this categorical dependent variable, the logistic regression gets used as a classifier. This
model maps variables between 1-0 (such as true or false and pass or fail). This two-outcome model is
also called binomial or binary logistic regression.
SUPERVISED VS. UNSUPERVISED LEARNINGCH. 2MACHINE LEARNING FOR MARKETERS

11
Unsupervised Machine Learning
As opposed to supervised learning, unsupervised learning involves only entering data for (x). In this
model, a correct answer doesn’t exist, and a “teacher” is not needed for the “learner.”
You’ll find two types of unsupervised machine learning: clustering and association.
Clustering – This type describes techniques that attempt to divide data into groups clustered close
together. An example of clustering is grouping customers by their purchasing behavior.
Association – This type describes techniques that create rules that explore the connections
among data. An example is helpful here: We might say that people who buy X product also often
buy Y product.
Semi-supervised Machine Learning
But wait … you’ll also find a third type of machine learning technique, semi-supervised machine
Learning. Think of this type as a hybrid between the two previously mentioned models. In most cases,
this type of learning happens when you’ve got a large data set for (x), but only some of (y) is definitive
and capable of being taught.
Semi-supervised machine learning can be used with regression and classification models, but you can
also used them to create predictions.
Classifiers
Decision trees build a series of branches from a root node, splitting nodes into branches based on the
“purity” of the resulting branches. You use decision trees to classify instances: One starts at the root of
the tree. By taking appropriate branches according to the attribute or question asked at each branch
node, one eventually comes to a leaf node. The label on that leaf node is the class for that instance.
This modeling is the most intuitive type of modeling. You’ve likely used some version of it in your
school or professional life.
Backed-up Error Estimate – In order to prune decision tree and keep the “purity” of each branch
of the tree, you must decide whether an estimated error in classification is greater if the branch is
present or pruned. A system to measure these issues takes the previously computed estimated errors
associated with the branch nodes, multiplies them by the estimated frequencies that the current
branch will classify data to each child node, and sums the resulting products. Training data gets used
to estimate instances that are classified as belonging to each child node. This sum is the backed-up
error estimate for the branch node.
Naive Bayes – Naive Bayes classifiers are based on applying Bayes’ theorem with naive independence
assumptions between the features. Bayesian inference focuses on the likelihood that something is
true given that something else is true. For example, if you’re given the height and weight of an indi-
vidual and are asked to determine whether that person is male or female, naive Bayes can help you to
make assumptions based on the data.
Reinforced Learning
As discussed in the previous chapter, reinforced learning is a machine learning evolution that involves
neural network development. Reinforced learning combines multiple layers of networks into complex
models that are useful for building intelligent models.

12
Neural Networks – This collection of artificial neurons (or machine learning and algorithmic layer)
gets linked by directed weighted connections. Each neural network layer feeds the next. The neural
network levels each have input units clamped to desired values.
Clamping – This action takes place when a layer, also called a “neuron,” in a neural network has its
value forcibly set and fixed to some externally provided value, otherwise called clamping.
Activation Function – This function describes the output behavior of a neuron. Most networks get
designed around the weighted sum of the inputs.
Asynchronous vs. Synchronous – Some neural networks have layers of computations that “fire” as
the same time, with their nodes sharing a common clock. These networks are synchronous.
Other networks have levels that fire at different times. Before you stop us, let’s clarify something for
you: We’re not saying that there’s not a pattern to how the levels handle input and output data; we’re
only saying that the levels aren’t firing in a precisely timed way.
Mathematical Concepts
Dimensionality Reduction – This concept uses linear algebra to find correlations across data sets.
Principal Component Analysis (PCA) – This process identifies uncorrelated variables called prin-
cipal components from a large set of data. The goal of principal component analysis is to explain the
maximum amount of variance with the fewest number of principal components. This process is often
used in utilities that work with large data sets.
Singular Value Decomposition (SVD) – This technique combines information from several vec-
tors and forms basis vectors that can explain most of the variances in the data.
Graph Analysis – This process involves using “edges” connected to other numerical data points to
analyze networks. The data points are known as nodes, and the edges are the ways in which they get
connected. Facebook’s EdgeRank, which was replaced by a more advanced machine learning algo-
rithm, got its name from graph theory.
Similarity Measures – Sometimes known as a similarity function, a similarity measure quantifies the
similarity between two objects. Cosine similarity is a commonly used similarity measure. This mea-
surement is used in information retrieval to score the similarity between documents. Clustering uses
similarity measures to determine the distance between data points. Shorter distances are equivalent
to greater similarity.

What Marketers Can
Accomplish with
Machine Learning
CHAPTER 3

14
Now that we know the language of machine learning, we’re ready to look
at specifically what marketers can do using machine learning. The ad tech
space is full of companies promising the next silver bullet for marketers.
Armed with your new knowledge of machine learning and related concepts,
we can begin to look past the veil toward what makes these tools, process,
and marketing services tick.
Marketing Automation
It’s highly unlikely that if you’re reading this guide you’ve not worked directly with the concept of
marketing automation. It’s even highly likely that you’ve played around with one of the industry’s
leading marketing automation platforms such as Marketo or HubSpot. Even niche tool providers like
MailChimp have created automation in their platforms to help marketers scale their efforts.
These automations often help marketers and sales teams streamline and manage tasks that human
hands once did. Drip emails are a great example. You can now build an email list, generate a number
of templates, and the system will email your recipients at the time and on the conditions you
instruct it to.
While these automations are highly valuable to marketers, the next iteration for these systems will be
layering in machine learning.
Using the above example of email drips, software providers like Mautic are already offering mail
automation that relies on a logic true. If your recipient Z takes action X, then it sends email Y. This
supervised learning system (remember that term from Chapter 2?) is a basic one.
The next evolution in such a system would come from Mautic learning how long it takes for recipient
Z to respond to emails and instructs your sales team on the best time to follow up an email with a
call based on an unopened email. Going even further, the system helps you to pick the best adjectives
and verbs for your subject lines based on previous subject lines and open rates. You’ve likely seen or
reviewed tools with these types of features, since they’re becoming more available and can greatly
impact the value of human capital working on marketing and sales initiatives.
Sending Frequency Optimization
Sending frequency optimization can also have a substantial impact on both your standard email mar-
keting initiatives and your drip campaigns.
Machine learning can help you answer the following questions:
1. How often should you pay attention to specific recipients and segment new
marketing messages?
2. How often should you follow up with leads?
3. What days are most effective for follow-ups and new marketing messages alike?
Before machine learning, marketers would leave frequency optimization up to testing and ROI anal-
ysis; however, frequency optimization was all but a measurement on full lists in most cases. Machine
learning allows marketers to carve lists into precise segments and to neatly personalize sending fre-
quencies for individual recipients.
WHAT MARKETERS CAN ACCOMPLISHCH. 3MACHINE LEARNING FOR MARKETERS

15
Content Marketing
How much time are you spending on administrative tasks, such as asset tagging, versus content cre-
ation? Tagging assets with relevant keywords is essential to making them searchable, but it’s also a
tedious, time-consuming task that most marketers would rather avoid.
Machine-learning technology can smartly include the most valuable or least expensive keywords in
copy. This technology can associate similar or related assets, making it easier to combine relevant
copy, images, and videos for target audiences. If your audience has consumed one bit of content, then
it can smartly recommend what to offer next for the highest conversions. Machine-learning technol-
ogy can also help predict what content will lead to the behaviors — sharing or engaging with content,
increased sales, and improved customer loyalties — you’re trying to gain from customers. Adobe’s
Smart Tag technology is available now to automate metadata insertion so that you can get better
search results while slashing the amount of time you spend on this task.
Each marketing channel presents a special set of requirements for the size and resolution of marketing
assets. When a new platform emerges — or if you decide to add a new channel — it could require the
time and cost of redesigning existing assets. For example, if you have a piece of content delivered to a
web channel or blog, machine learning can smartly crop that content for a mobile channel or reduce
the copy in smart ways. With machine learning, you can shorten visual content and videos to optimize
experiences for different channels based on the patterns by which people are consuming them.
Machine learning will either offer recommendations — or provide a first draft of the new content
— that can then help accelerate the pace by which you get those different pieces of copy, creative
graphics, or videos published to various channels and distributed to the selected audiences.
You don’t want to have to create massive volumes of content and hope that only some of it will be
effective. What’s important is being able to create the right content that’s effective in your channels,
learn from that content creation, and then develop more content based on those insights as you
expand from that point.
Machine learning can give you the intelligence needed to quickly determine what’s working as well as
recommend what’s needed to amplify your strategies that might better connect with your audience.
The learning part of machine learning means that, over time, the machine becomes smarter. We’re
still in the early stages with this knowledge evolution, but machines could potentially learn so quickly
that you could remix, reuse, and adapt content almost instantaneously, test the content you’ve cre-
ated, and learn whether what you’ve created will be an improvement over your previous campaign or
whether you need a different approach.
Ad Platforms
One of the first platforms to incorporate machine learning into systems and processes that marketers
use every day are media-buying software sets. As an example, Google’s quality score helps determine
which ads are most relevant to searchers and the cost per click of ads by using factors such as click-
through rate, ad copy relevance, and landing page quality.
Programmatic Display
With the growth of mobile internet consumption, banner ads and other display advertising have had
to undergo a major change in order to retain their value. Further, with the rise of better and better

16
technology, marketers are looking for ways to segment their message in a more precise manner and
not broadcast to a large audience in hopes of motivating a few people.
Programmatic advertising allows marketers to take matters one step further by measuring and
pivoting on advertising strategies in near real time. Where big data has allowed for segmentation, pro-
grammatic advertising has allowed for marketers and advertisers to see how these segments perform
on specific ads at specific times on specific devices, for example. This type of advertising also allows
for more control over spend and higher ROI on display ads.
AdWords Scripting
Google has developed much great functionality in its system for ad control and information in managing
ads. Its resources will help you find new segments that you should be reaching and give you information
on underperforming ads, groups, and keywords. However, tasks and processes are available to help an
AdWords practitioner in ways that AdWords doesn’t offer: AdWords Scripts fills this need.
Adwords Scripts runs within the Google AdWords platform. The option can be found in the campaign
navigation under “Bulk Operations > Scripting.” These Scripts get written through Javascript, although
you can find many premade Scripts, too. These Scripts often work as a form of supervised machine
learning in many cases. AdWords advertisers specify the input and the output determines the func-
tion. An algorithm connects the input to the output .
Predictive Analytics
Predictive analytics are used in many verticals and professions. Essentially, these analytics are models that
help find and exploit patterns in data. These patterns can then be used to analyze risks and opportunities.
In the marketing world, predictive analytics layers on top of our standard analytics data to give us
more insight and help shape marketing actions. Previously, we’d look at raw data like “sessions” that
would give us insight into our analysis of ROI based on base metrics of lifetime value for a session.
Now, we can predict with more precision the exact value of each individual session based on their
onsite actions and the sources of their referrals. As with programmatic advertising, we’re able to
interact with the potential customers represented by those sessions in highly tailored ways, including
chatbots, which we’ll discuss in more detail in a later chapter.
Customer Churn
Keeping customers is as pivotal to growth as getting new customers. Analytics can help understand
behaviors that lead to customer churn and help marketers craft strategies to reverse churn. Predicting
customer churn is a valuable piece of this puzzle. Through machine learning, behaviors of past users
that are no longer customers can give us a data set of actions to apply to the current customer base
to see which customers are at risk of jumping ship on you.
Most predictive churn models have their bases in logistic regression and binary models (aren’t you
glad you read through all those definitions now?). Micro-segmentation and the use of machine learn-
ing can help marketers and sales teams understand when a customer or client may be ready to jump
ship and help spark an action that can keep churn numbers low.
Computer Vision
Computer vision is exactly what the term sounds like — it’s how machines “see.”
Machine learning, especially reinforced learning, has allowed machines to have ever-increasing capa-
bilities related to image recognition. A great everyday example is the facial recognition that Facebook
uses when suggesting people to tag in a photo. Facebook’s facial recognition technology has “learned”

17
the faces of users over time as they’ve been tagged by the endless amount of photos that make their
way into the Facebook universe.
Computer vision has practical uses in marketing.
For example, Sentient Aware offers software that lets its users serve customers with products that are
visually similar to the products that they choose. This style of product serving could have a significant
benefit over the traditional use of tagging, especially when dealing with new customers whose buying
habits are not yet known.
Snapchat and Instagram have made visual-based social listening increasingly important. Ditto and
GumGum provide social listening tools that can enhance reputation-management efforts. For exam-
ple, brand marketing folks can receive alerts to tell them when their company’s logo appears in a
meme or image that might need their attention.
Segment Comparison
Audience segmentation has always been an important part of advertising. Knowing the members of
your audience and where they’re coming from offers marketers incredibly valuable information. Until
the invention of the internet, gaining that data was never done in real time. Now marketers can gain
almost real-time access to the demographic-based data of their consumers and create actions that
interact with them.
Only a decade ago, marketers rejoiced when they gained access to data such as age, sex, location, and
the length of time that users interacted with their messages. Now marketers can create micro-seg-
mentations as well as measure and compare how each segment reacts to different messages.
Google Analytics offers behavioral-based demographic data such as affinity groups for a user.
Companies such as Jumpshot offer extremely detailed segmentation, offering the ability to let you
know which micro-segments purchase from your competition, at what times, and how they’re finding
them. Furthermore, these tools can tell you which micro-segments buy a $1,000 watch and which buy
a $100 watch, which can give you a better analysis of where to spend the dollars in your own market-
ing and advertising budget.

Successful Machine
Learning use Cases
CHAPTER 4

19
Now that we’ve pushed through both generalities of machine learning, its
basic concepts, and how they apply to areas of marketing, it’s time to dive
into the specifics of how companies are using these processes.
We’ll look at a number of examples of how major companies are using machine learning to interact and
deliver content to customers in an increasingly targeted and personalized way. We’ll also look at how
some marketing companies are using this technology to increase their insights and their clients’ ROI.
Target Identifies Pregnant Customer Before Family Does
This use case is one that Target’s PR department probably didn’t cook up. The story went viral and was
the first experience that many people outside the marketing and advertising spaces had with the way
major retailers were using machine learning to target their customers.
Target assigns all their customers Guest ID numbers tied to their credit card, name, and email address.
Further, they use customer rewards products such as their Red Card and Cartwheel to gain even more
data points to tie to these Guest IDs.
In this case, Target used the data from women who had signed up for their baby registry. They looked
at the historical purchases to see what the data profile for a pregnant customer looked like. Going
back to our machine learning terminology, they likely used a variation of a Naive Bayes to create a
classification for the group.
Next, Target sent this classified group coupons for baby items. The recipients included a teenage girl
whose father was unaware of her pregnancy. The coupon likely wasn’t the way the girl wanted her
father to find out about her expected child or the way Target wanted the world to find out about its
customer targeting. However, this story is an interesting and practical example revealing how retailers
are collecting, learning about, and using data points.
Adobe Knows What Your Customers Want and When They Want It
Adobe’s Marketing Cloud puts several tools in the hands of marketers that can help them immediately
put machine learning to work.
The Adobe Audience Manager allows users to centralize all of their customers’ data into a data man-
agement platform. From this platform, they can create distinct customer profiles and unified audience
segments. Understanding audiences and segments allows Adobe’s customers to tailor content specif-
ically for these consumers in their Experience Cloud products. Beyond delivering a targeted message,
the tools in Adobe’s products allow you to find the segments in your customer base that have the
highest yield and focus your effort and spend on those targets, thus maximizing your ROI.
Adobe has also brought about Virtual Analyst, powered by an artificial intelligence (AI) interface with
its Adobe Sensei product. Virtual Analyst continually processes data and uses predictive algorithms
and machine learning to drill into specifics of your business operations. Virtual Analyst is like the real-
life data scientist working with you in your company.
Adobe reports several benefits to customers using Virtual Analyst: increased revenues, major cost
savings, mitigated risks, and bug detection.
Facebook Filters Posts and Advertisements for Users
Facebook wanted to understand more about the text content its users were sharing so the social
media giant built DeepText. This reinforced learning platform helps users to make sense of the context
of content and allows Facebook to filter spam and bring quality content to the surface. Of course,
DeepText also has implications on ad segmentation and delivery.
SUCCESSFUL MACHINE LEARNING USE CASESCH. 4MACHINE LEARNING FOR MARKETERS

20
The DeepText system is a deep neural network that uses FBLearner Flow for model training. The
trained models go into the FBLearner Predictor platform. The scalable model allows constant model
iterations for DeepText.
To understand the scale of DeepText, keep in mind that Facebook has 1.86 billion monthly active users.
This user base is worldwide, and thus DeepText must have an understanding of many languages. The
level of understanding goes beyond basic Natural Language Processing. Instead, the goal of DeepText
is to have an understanding of the context and intent of the content, not simply what the content
says only.
DeepText achieves contextual understanding by using a mathematical concept called “word embed-
dings.” This practice preserves the semantic relationship among various words. With this model,
DeepText can see that “bae” (a term of endearment) and girlfriend are in a close space. Word embed-
dings also allows for terms across languages to have similarities aligned.
One practical use of DeepText can be found in the Messenger application. The Messenger app
uses AI to figure out if someone is messaging someone with the intent to get a ride from a taxi
somewhere. If Messenger recognizes this request, it offers up the options to “request a ride” from
a ride-sharing application.
Future use cases that Facebook has floated include the ability for the system to understand when
users want to sell products they’ve posted and offer them tools to help with the transaction. DeepText
could also be help for bringing to the surface high-quality comments on posts by celebrities and other
large threads.
With it’s new focus on improving the news feed, you can see where Facebook could deploy DeepText
to help flag issues such as fake news at scale.
Additionally, Facebook is using AI to create chatbots that talk like humans. To do so, the social net-
working giant has developed a free software toolkit that people can download to help Facebook
compile data, view research, and interact with others participating in the projects. The objective is to
get computers to understand human conversations without failing; to do that, individuals can provide
computers with real conversations and language used online to teach them. For example, information
from Reddit discussions about films can train computers to converse with humans about movies.
Clarifai Identifies People and Objects in Videos for Easier Media Indexing
Clarifai is an image and video recognition company based in New York. The firm has raised $40 million
over two rounds from Menlo and Union Square Ventures to build a system that helps its customers
detect near-duplicate images in large uncategorized repositories.
Clarifai’s technology can be used through one of the e-commerce examples we discussed in Chapter 3.
The company’s “Apparel” model, which is in beta, claims to be able to recognize over 100 fashion-re-
lated concepts including clothing and accessories. Further, the company’s “Logo” model can be used
by companies looking for a social listening tool for Snapchat and Instagram.
Clarifai’s “Demographics” model offers an interesting opportunity for marketers as well. The company
is able to analyze images and return information on the age, gender, and multicultural appearance for
each face detected. The opportunities for advertisers to target specific demographics in an increas-
ingly visual internet landscape , where there is less text content to inform the intent, are arriving at
the perfect time.
And just when you thought what Clarifai was doing was cool in itself, the company also allows custom-
ers to “train” their own model.

21
Sailthru Helps Customers Develop Digital Marketing Campaigns
With Better ROI
Sailthru is another New York–based company building solutions around machine learning. Their
solution focuses around promotional emails. Sailthru’s system learns about customers’ interests and
buying habits and generates a classification based on the information. This classifier tells the system
how to craft the message and when to send the message to optimize performance.
One Sailthru customer, The Clymb, saw a 12 percent increase in email revenue within 90 days of
launching the Sailthru personalization model.
Sailthru also offers prediction and retention models that allow companies to be able to anticipate
customers’ actions. After turning on these models, The Clymb saw a 175 percent increase in revenue
per thousand emails sent and a 72 percent reduction in customer churn.
Netflix Personalizes Video Recommendations
Netflix is incredibly transparent about how it has used machine learning to optimize its business. The
most obvious place that you can see the effects of their use of machine learning is on the Netflix
homepage. Netflix personalizes its homepage for every user.
One of the issues Netflix faces with its homepage is the sheer amount of content and products
from which it chooses. The homepage has to not only pull content that the user is likely to want to
the surface, but it also has to serve that content as a doorway to other content that the user may
find interesting.
Netflix uses a series of videos grouped in rows. The company can group these rows by basic metadata
such as genre. However, Netflix also creates rows based on personalized information, such as TV
shows, that are similar to other shows the user has watched.
This system that Netflix uses is a great example of graph analysis. Netflix is able to examine the connec-
tion between various data points and recommend content based on the “edges” between the points.
As an example, “The Rock” could be one edge that connects “The Fast and the Furious” movies with
“The Scorpion King.”
Concepts such as Naive Bayes and logistic regression are likely used to help create profiles and
match those profiles with the outputs from the content graph analysis. Naive Bayes would help clas-
sify groups of users by their behaviors, and the logistic regression would qualify whether that group
should be served each type of content. For example, if a user is watching mostly content focused on
preschoolers, Netflix knows not to serve horror movies as a suggestion.
On top of these two types of machine learning, with independent focus points on content and user
behavior, Netflix also conducts A/B tests on the layout of the homepage itself. Historically, Netflix had
displayed rows such as “Popular on Netflix,” but the company realized that this arrangement was not
personalized for every user. In our example above of the preschool Netflix user, getting recommen-
dations for the new season of “Orange Is the New Black” because the show is popular doesn’t make
much sense for this audience. However, showing this user content grouped in rows by their favorite
children’s entertainment characters would make logical sense.

Identifying who
is truly leading
the field in
terms of success
in SEO is still
not an exact
science.
SUCCESSFUL MACHINE LEARNING USE CASES

23
iPullRank Predicts Visibility of Sites in Organic Search in Vector Report
Identifying who is truly leading the field in terms of success in SEO is still not an exact science. While
you can easily view something such as a content marketing campaign and grade a company on its
efforts in this area, SEO by its nature is more like a black box.
As such, iPullRank set out to create a view of the Inc 500 companies and their overall and market-level
organic search performance. This study intended to predict a site’s performance in SEO from relevant
link factors from Ahrefs, cognitiveSEO, Majestic, and Moz.
The data collection involved the following:
1. Scraping the Inc 500 list from the company’s site
2. Pooling the 100 winners and 100 losers from Searchmetrics
3. Establishing the URLs from each company
4. Pulling all the available domain-level metrics from Moz, SEMrush, Searchmetrics,
Majestic, and Ahrefs.
5. Placing the 700 domains into cognitiveSEO
6. Using cognitiveSEO’s Unnatural Links Classifier
Next, iPullRank reviewed 116 features for every domain and used logistic regression, cross validation,
random forest, and lasso to analyze the data.
From here, we collected two sets of training data. We used the 100 SEO winners and 100 SEO losers
from Searchmetrics for our 2014 and 2015 Inc 500 data pulls, respectively. We then used the training
data from 2014 to select the model. With the 2015 training data, we updated coefficients with the
chosen model.
The sample size for 2014 training data was relatively small, and we knew that any statistical models
would be sensitive to the data points selected as the training data set. We used 5-fold cross validation
to choose the model, solve the sample size data, and reduce variance of the model.
We used several machine learning methods to select variables, including random forest and lasso. In
the end, we selected eight factors in the final logistic regression model which are believed to influence
the overall performance of a webpage. These eight important factors are as follows:
SUCCESSFUL MACHINE LEARNING USE CASESCH. 4
Variable Name Metric Provider Description
umrr Moz The raw (zero to one, linearly-scaled) MozRank of the target URL
refclass_c Ahrefs Number of referring class C networks that link to the target
CitationFlow Majestic A score helps to measure the link equity or “power” the website or
links carries
TopicalTrustFlow_Value_2 Majestic The number shows the relative influence of a web page, subdomain
or root domain in any given topic or category
Rank SEMrush Site rank on Google
Unnatural.Links CognitiveSEO Number of unnatural links
OK.Links CognitiveSEO Number of OK links
fspsc Moz Spam score for the page’s subdomain
MACHINE LEARNING FOR MARKETERS

24
Before predicting winners and losers for the 2015 Inc 500 companies, we used median imputation
to handle missing values. Then we updated the final model coefficients with 2015 training data. The
prediction results are plotted below:
We noticed that most companies registered predictions as losers (probability <0.5) based on these
plot points. This result follows from the huge difference in the distribution of the 2015 training data
set and 2015 Inc 500 company data. Put simply, the SEO winners and losers we used for training data
are often more advanced in their SEO strategies and techniques than the standard Inc 500 company.
Therefore, we came up with the alternative to grade each company based on the relative performance
within Inc 500 companies.
We measured the overall performance in the following two ways:
1. weighted average score of each variable
2. the predicted probability of being a SEO winner
The lower score between the above two methods was used to better identify underperforming com-
panies from an SEO perspective. Overall performance using predicted probability of being a SEO
winner gets decided by the following rules:
A Top 10%
B Top 10% – 25%
C Top 25% – 45%
D Top 45% – 65%
F Top 65%

25
We found that 2015 Inc 500 companies had better organic search performance than 2014 Inc 500
companies in terms of spam score, trust metrics, link popularity, visibility, and unnatural links. In gen-
eral, however, most Inc 500 companies have high grades in spam scores and unnatural links, but have
poor link popularity and low trust metrics. This finding shows a general lack of off-site SEO and con-
tent marketing strategy by the group as a whole.
Below are the number of companies being assigned to each grade for 2014 and 2015, respectively:
Grade 2014 Inc. 500 2015 Inc. 500
A 34 47
B 63 76
C 54 102
D 127 91
F 222 183

How Machine
Learning Guides SEO
CHAPTER 5

27
Our Vector Report served as a delicious appetizer for the world of machine
learning and how it applies to SEO and search. As we’ve noted, the core of
modern search builds upon models based on machine learning. As the field
has expanded, both search engines and search practitioners have worked to
incorporate more robust machine learning processes into their methodolo-
gies and technologies.
Google and Machine Learning
Most of our conversation will center around Google and its expansion of machine learning technolo-
gies and where they might be going next.
Ranking Algorithm and Beyond
Remember when we talked about the original PageRank algorithm and how this algorithm can be
treated as unsupervised learning? We could make arguments against the early Google algorithm being
what we consider machine learning today, but the reality is that the algorithm aimed to make sense of
patterns in unstructured data around the web and put those patterns into a coherent output — we
know them as SERPs. Detractors from this view will point out that the algorithm is only one part of the
overall information retrieval structure and not by definition pure machine learning.
What can’t be argued is that Quality Score — arguably the economic engine of Google — is a
machine learning–based program. This score uses a large amount of signals to predict future out-
comes. The largest signal in this mix is click-through rate (CTR) for well-established accounts with
history. A myriad of other signals make up the remaining Quality Score mix, but based on the effi-
ciency of Quality Score and its staying power, SEO sleuths could see the writing on the wall that
Google would eventually take a “smarter” approach to its overall algorithm.
RankBrain
Enter RankBrain. RankBrain is an algorithm-learning AI that Google verified existed in October 2015.
RankBrain isn’t the algorithm but part of an algorithm, one that Google has stated is the third most
important factor today behind links and content.
RankBrain’s real power is understanding never-before-seen queries and delivering results based on its
understanding of the term. RankBrain groups search queries close to each other in linguistic similari-
ties into “distributed representations.” Put simply, RankBrain is trying to predict what people “mean”
in their search queries by looking at their intent.
Google’s RankBrain can “self-learn” and form new word associations over time. This ability makes it
capable of working faster than a human “training” the program to recognize linguistic similarities.
The principles of RankBrain are similar to the word2vec tool, so while RankBrain is a bit of a closed
box outside of what Google has let out, we can get a better idea of its structure from looking at this
comparative AI tool.
As an open-source toolkit, word2vec uses a text corpus to count the distance between words and
product vector representations of words and phrases. The closer in distance the vectors are, the more
similar in meaning they are.
According to Google, no one way exists to optimize content for RankBrain specifically. Further, the
meat and potatoes of RankBrain’s effects will be found in the longer tail terms where you’re less than
likely to be fishing. The main intent for RankBrain is to help with brand-new queries.
HOW MACHINE LEARNING GUIDES SEOCH. 5MACHINE LEARNING FOR MARKETERS

28
However, the one place that RankBrain has the largest effect is in content creation. The idea of
keyword density having a major effect on search died long ago, but RankBrain is an overall death
knell, putting the focus on comprehensive content that provides immense value to readers. This
kind of overwhelmingly informative piece of content on a subject is what we’re creating with this
guide as an example.
Google Patents and the Future
With RankBrain launched, and the value that Quality Score plays in its advertising system, Google is
clearly moving quickly toward innovation in machine learning to keep its market lead in the world of
search. The next step for SEO practitioners is to look at what the company’s patents tell us about its
future intentions about the concept.
Methods and Systems for Classifying Data Using Hierarchical Taxonomy
Filed in 2012, this patent for classifying data using a hierarchical taxonomy is a method and
system for classifying documents. The use case to which Bill Slawski attributed the model at Google
is customer support, a system that helps to classify customer issues and route them to the correct
channels. In this use case, Google would group various issues related to each other. For example, pay-
ment processing would be a subcategory under which various issues would be classified. From this
level of classification, the system would build a hierarchy of importance of the issues (also discussed
as documents).
This same concept could be applied to ranking algorithms, and the patent gives us insight into how its
ranking module handles ranking algorithms in the use case of customer service:
In some implementations, the ranking module ranks document clusters according to one or
more metrics. For example, the ranking module may rank the clusters according to the document
quantity in each cluster. A cluster with many documents may represent a relatively significant topic,
such as a product issue.
Need another example? The ranking module may rank the clusters according to an estimated
time to resolution of an issue represented by the cluster. Issues represented by a cluster
“software update,” for example, may typically be resolved faster than issues represented by a cluster
“hardware malfunction,” a label assigned to the cluster, a number of documents in a cluster, a desig-
nated importance of subject matter associated with a cluster, identities of authors of documents in a
cluster, or a number of people who viewed documents in a cluster.
One more example for you: A cluster that represents an issue that historically has taken a longer
time to resolve may be ranked higher than a cluster representing an issue with a shorter historical
time to resolution.
This patent shows that Google has done internal work to classify and rank content based on learned
value of documents. Bounce rates and click-through rates (CTR) could take the place of “estimated
time to resolution” as data points that quickly make this effective addition to ranking algorithm.
Topics and content that matches what the system perceives to be high-quality content based on past
user-based signals could float to the top of the results. In many ways, Google could use this type of
document-based understanding as a citation-based replacement to links. Brands and websites get
value based on how much internet users talk about them online.

29
Since the system is constantly “learning” from its generated outputs and adding new “training” from
these learning experiences, a reinforced learning system like this one, focused on organic search,
would simply need time to learn to be efficient enough to be added to the ranking algorithm.
Searchmetrics Content Experience
Since Google has made machine learning a priority, organic search tool providers must do the same to
keep their customers armed with the right technology for the changing market.
Searchmetrics created a tool called “The Searchmetrics Content Experience.” This tool uses rein-
forced learning to help SEO managers and content creators within an organization collaborate and
create better content.
In this tool set, marketers are able to research topics, optimize content, and measure results.
SearchmetricsusesreinforcedlearningtodeliverrelatedtopicsinitsTopicExplorertool.Searchmetrics
uses NLP to create a group of nodes and edges around related terms.
The optimization tools in the tool set deliver help and support to writers through interactive param-
eters and live content scoring. Writers learn in real time how to create the most effective piece of
content possible. These efforts are all based on reinforced learning around 250 billion continually
updated data points in the Searchmetrics index.
What Searchmetrics has effectively created is a tool that has SEO managers thinking, collaborating,
and creating content around concepts that are similar to what RankBrain is doing to answer questions
for new queries.
The Future for SEO
Based on the rise of RankBrain, the feelers we’ve gotten from Google on future machine learning
plans and the tools being created in the SEO space makes one point clear: Comprehensive content is
future-proof SEO.
While Panda killed off the days of automated content at scale — offering quick wins for SEO managers
looking to get ROI without expelling much energy — the reality is that lower quality content produc-
tion is still happening.
Content is still commoditized, and getting content at scale is difficult and expensive. Great content
creates great links, which have real impact on today’s algorithm — the number-one factor on search
rankings today. Great content also yields value with systems such as RankBrain that are likely to sur-
face quite comprehensive articles for intensely specific queries.
Based on the other concepts we looked at in this chapter, content that garners quality signals from
search visitors is also likely to have an increasing impact in the future, especially as Google is able
to associate those data points with content, learn the meaning of that corollary data, and apply it
to ranking results.

Chatbots: The
Machine Learning
you are Already
Interacting with
CHAPTER 6

31
Chatbots are probably the most common format of reinforced learning that
people are interacting with daily. Since chatbots can pass the Turing test,
some people probably don’t know that they are interacting with chatbots in
some cases.
The most widely used versions of chatbots today are likely the AI assistants from the major tech com-
panies such as Siri from Apple and Alexa from Amazon. While chatbots differ widely in how advanced
they are and how they’re built, most chatbots have some basic similarities.
Platform
Chatbots are usually hosted within a messaging platform of some type. If you use Slack, the chatbot
that comes bundled with your account from the outset is the Slackbot. This bot can help with many
Slack-based activities and reporting.
Another example that we already looked at is the way Facebook uses DeepText to interact with its
users in the Messenger app.
Language Processing
This element is a vital layer of any chatbot. Human language communication between man and
machine is impossible without a Natural Language Processor (NLP).
Essentially, an NLP refers to components that allow chatbots to carry out a conversation that’s nearly
identical to a person-to-person human conversation. NLPs use deep learning to not only study human
input but also generate a human response. According to Chatbots Magazine, newer, smarter chatbots
using NLP should be able to carry out the following functions:
• Summarizing large quantities of text into concise explanations
• Responding to open or closed questions (“Is Washington, D.C., the capital of the United
States” compared to “Why is Washington, D.C., the capital of the United States?”)
• Inferring conference clues (that is, relating an object with words in a sentence)
• Ambiguity (sorting out the context and meanings of words in sentences)
• Morphology (the ability to separate individual words into morphemes, the smallest unit
of meaning for words)
• Semantics (the meaning of sentences or words in natural human language)
• Text structure (the usage of punctuation, spacing, and text)
• Sentiment (the emotional “feeling” conveyed by the speech patterning, for example,
“happy” or “sad”)
Amazon powers Alexa with its own NLP called Amazon Lex. Most chatbot builders use third-party
NLPs rather than building their own, and Amazon has made Lex available to developers.
Application
This stage is where the machine learning happens. Despite the advanced options we’ve shared with
you, chatbots can run on simple supervised machine learning models. For example, you can train
a chatbot to respond to various questions from customers with predetermined responses. As new
questions come in from customers, you can classify them within various categories of responses, thus
“training” the application. This part of bot development can now find a third-party home at services
CHATBOTSCH. 6MACHINE LEARNING FOR MARKETERS

32
such as Amazon’s Machine Learning Service. When the process being used to manage the appli-
cation moves into the neural network level of computing, the chatbot moves into the world of AI.
Database
The database is the arrival location for the data points that the application computes. In the example
in Chapter 5 with our customer service chatbot, this database can be as simple as the list of questions
and responses. Obviously, the more advanced a chatbot is, the more data and the more advanced the
machine learning that deals with that data will need to be.
As you can see, chatbots are a key application that connects the value of machine learning with tasks
people do every day.
Now that we understand the basic components of a chatbot, let’s look at some use cases for the
technology.
WeChat Empowers Its Users
WeChat is a social media application brought to life by Chinese company Tencent, the company
behind the QQ desktop messaging software. The app has more than 899 million active users spending
over an hour a day on the app on average. WeChat is huge in China, but it has yet to break through
the social media realm internationally. A special projects team within Tencent created the product.
WeChat runs contrary to Facebook in many ways. Where Facebook launched Messenger to “unbun-
dle” services that users may want to use, WeChat gives you everything you could want in a mobile
experience in one application. In other words, WeChat is not only a social media home for many users,
but it’s also the method for shopping and payment services.
WeChat has built an API for developers to build chatbots on their applications. Today, this API lets
users do many interesting activities such as:
• get a ride share
• order food
• buy movie tickets
• order from the nearest Starbucks
• track your daily fitness progress
• book appointments
• pay bills
In many ways, WeChat served as the beta test for chatbots in the social media space for the U.S.
market. Major technology companies in the West took notice, and Windows developed a chatbot
called Xiaoice. This NLP application from Microsoft has more than 40 million users in China and Japan.
Xiaoice is capable of leading long conversations with distinct personalities. Microsoft mined the
internet for human conversations on Chinese social media and used reinforced learning to develop a
chatbot that can adapt its phrasing based on the positive or negative cues given by the user.

Chatbots
are the most
common format
of reinforced
learning that
people are
interacting
with daily.
CHATBOTS

34
CNN’s Bots Share the Latest News
In 2016, CNN began rolling out a chatbot to interact with younger demographics on social media plat-
forms. The first bots rolled out on Messenger and Kik, and they were created to give users the latest
news in a conversational format that mirrors their everyday messaging application use.
CNN specifically launched the Kik application with an
interactive explainer to help users understand the devel-
oping news better. Additionally, CNN launched bots on
LINE and for Amazon Echo.
Commenting on the news organization’s view on its suc-
cess, Alex Wellen, CNN’s senior vice president and chief
product officer, told The Drum:
“We’ve seen the most growth on LINE, where
our chatbot publishes a handful of international
stories each week; we see high engagement on
Kik, where our chatbots invite the audience to
experience the news in a ‘choose-your-own-
adventure’ format; and we’ve observed the
most experimentation on Facebook Messenger
and Amazon Alexa, where people can ask and
receive responses to open-ended questions
about the news. CNN defines success metrics
for each platform, from reach to engagement
to monetization, with the overarching goal to
learn, scale up, or wind down the product.”
We’re still in the early days of chatbots. Companies are still trying to understand how to quantify the
value of the interactions they’re having on the various applications.
Wellen gave a peek at the possible future for CNN’s chatbots: “The next frontier is real-time multime-
dia chat bots that return intelligent audio or video responses across mobile, IoT (Internet of Things),
and connected TVs.”
Magic’s Chatbot Offers Complete Virtual Assistant Capabilities
Magic is an online virtual assistant service that launched in 2015. This service has created chatbots in
several applications, including Slack, Messenger, and through SMS (Short Message Service). We might
think about Magic as an example of conversational commerce, an important subset of chatbots.
We’ve already looked at a few examples of conversational commerce with the examples of WeChat and
Messenger allowing you to order a ride share. Magic allows you to speak with a virtual assistant to help
you with dozens of tasks that an administrative assistant would handle, all through the applications you
use every day. These virtual assistants save data points that allow them to iterate the tasks later.
We know you might be thinking, but we’re going to tell you anyway: Magic is not a chatbot. While
Magic is an interesting example of how companies can enter the conversational commerce space, live
human operators manage the tasks in Magic. You won’t find any AI or reinforced learning happening
as of the date of this writing.

35
However, with the capabilities of companies to leverage an ever-expanding world of third-party
reinforced learning tools, the interactions happening at Magic today could serve as the training for
tomorrow’s AI-based virtual assistant.
We’re still much entrenched in the early days of chatbots. This time period feels like the early days of
mobile applications, and developers should be salivating at the untapped opportunities.
While the idea of developing a conversational bot may seem like a cumbersome task, remember that
you can start small. A chatbot simply needs to be built on the four components we discussed earlier.
A conversational commerce–based chatbot is possible with a human-generated database and basic
training over a supervised machine learning model. The tools and infrastructure are becoming more
readily available for all developers. With the increasing exposure of developers to these tools, we’re
likely to see an explosion of chatbots and related environments and economies over the coming years.

How to Set Up
a Chatbot
CHAPTER 7

37
In Chapter 6, we discussed the basics of chatbots and how they are import-
ant to the growth of machine learning. Now, we’ll dive into the nuts and
bolts of how you can set up a chatbot yourself. We briefly touched on the
overall structure and components of a chatbot, as well as some of the third-
party tools that exist to create chatbots, but we want to dive deeper into the
concept and give you a strong idea of exactly how you can put down this
guide and get to building your first chatbot.
Creating a Facebook Chatbot
Creating a Facebook chatbot from scratch isn’t extremely difficult, but the process can be a technical
one. Below, we present two versions of this process: one for individuals who don’t have a high degree
of coding knowledge and one for those readers who are more technically proficient.
Chatfuel
For those of you who don’t have a technical background or access to a developer, Chatfuel may be
the answer for you.
Backed by Yandex, Russian entrepreneurs Dmitrii Dumik and Artem Ptashnik founded . Chatfuel in
2015. Chatfuel builds chatbots on the Facebook Messenger platform. Users can train their own bots
in the Chatfuel system. Natural Language Processing (NLP) in the system allows training information
from the user to be applied to a larger web of related terms and phrases.
Chatfuel walks you through the same process that developers use as they code their applications.
Once you sign up via your Facebook account, you’ll choose a page that you want to associate with your
application. You may need to create a page if you intend to test your application before launching it.
HOW TO SET UP A CHATBOTCH. 7MACHINE LEARNING FOR MARKETERS

38
Blocks
The first and most basic setup for the Chatfuel bots is to create and link “blocks.” Blocks are essentially
messages sent to users. For example, below is a “welcome” block in the back end of the system:
And what follows is how the block looks in Messenger.

39
The “link” here is the “Main Menu” button. When you press it, Chatfuel gives you the following:
You’ll remember that this choose-your-own-adventure style of logic is similar to some of the function-
ality CNN built into its Kik-based news bot.
Chatfuel AI
You’ll get the most value from using the Chatfuel-integrated AI. This section of Chatfuel is the place
where you’ll train your chatbot how to handle questions. You can respond with basic text, images,
video, and other options.
Payments
Chatfuel offers a payments option that makes conversational commerce an option for bot builders.
Connecting to Facebook
The first step in creating your own Facebook chatbot through development is to create a working
endpoint that delivers a 200 response code. Put simply, this step lets you talk directly with Facebook.
Hartley Brody does a great job walking you through this step with a Github project to help you every
step of the way with code. To execute this process, you’ll need Heroku, which is basically a cloud plat-
form that allows companies to essentially become their own app companies by producing, sharing,
tracking, and scaling their apps. Heroku offers app developers the advantage of avoiding complicated
infrastructure hardware or scaled servers.
Choosing the Right Facebook Page
You’ll see that this is the second basic step to create your chatbot through third-party applications as
well. You can use a Facebook page you currently have or begin a new page, but think of this page as the
identity of your bot. It’ll be the group of people you can broadcast your bot to, and it’ll be the entity
with whom your users will interact. Be careful with your choice and testing.
If you don’t already have one, you need to create a Facebook page. The Facebook Page is the “iden-
tity” of your bot, including the name and image that appears when someone chats with it inside
Facebook Messenger.

40
If you’re creating a dummy one for your chatbot, it doesn’t matter what you name it or how you cate-
gorize it. You can skip through most of the setup steps.
In order to communicate with your bot, people will need to go through your page, which we’ll look
at in a bit.
Create Your Facebook Application
Now you are ready to create a Facebook App for your bot. You’ll need to name your bot, categorize
it, and give it some other basic info.
After you’ve created your App ID, you’ll have the option of selecting the product you want when set-
ting up your app. You’ll choose Messenger, which serves as the platform for your chatbot.
Set Up the Messenger Application
All of the steps below need to happen in Facebook before you get into the code portion of your
chatbot.
First, create a Page Access Token by clicking through the authentication flow (see below).
You’ll use this token to authenticate your requests when you try to send a message or reply to
someone.
You’ll then need to create a webhook. Facebook will ask you for a callback URL, verification token,
and subscription hooks. Hartley Brody’s tutorial and Github repository have more information on the

41
specifics of these items. The Facebook Developers portal also has vital information on the Messenger
Platform that can help you.
After you’ve configured your webhook, you’ll need to subscribe to the specific page for which you
want to receive message notifications.
Get Chatty With Your Bot
Start sending yourself messages through your newly connected application. Go to the Facebook page,
and under the options button next to “Share,” select “View Page as Visitor.”

42
This selection will allow you to message the page as a visitor.
Most developers use Heroku when building Facebook apps because Heroku allows developers to
work with the language that’s most natural for them to use when coding.
Beyond Hartley Brody’s framework, you’ll find a few others worth noting that may help you or your
developers with some plug-and-play code to get your chatbot talking.
• Messenger Bot Tutorial
• Charca bootbot
• Facebook Chat bot with Node and Heroku
One commonality you’ll notice is the way these chatbots work: They classify anticipated user ques-
tions or keywords and match them with responses.
Submit Your App for Review
Before Facebook verifies your application, only you and your page admins can test or launch the chat-
bot you’re creating. Facebook has to review all apps, but chatbots are especially large security risks.
The capabilities to spoof human behavior, send malicious links, and use processes to retrieve private
data are too real with this technology.
While you’re testing your bot, only you and other page admins can message the bot directly. You’ll
have to go through a review process before your bot is open to the world, ready to chat with anyone.
Facebook seems to be quite thorough in its review process, and with good reason. The code for a
messaging bot runs on your own servers and could change at any time, without Facebook knowing
it’s changing.
As long as your application is a legitimate chatbot set up to enrich the user experience of people who
like your page, you shouldn’t have any issues.
Botsify
Similar to Chatfuel, Botsify allows you to build a chatbot on Facebook Messenger free. It also offers a
paid version of its system to create a chatbot for a website. One of the interesting parts that’s special
about Botsify is the ability for bot builders to use what it calls “stories.”
Botsify Stories
Stories are essentially user stories, but they also serve as the real training data for the bot. The creator
builds out an entire possible interaction or series of interactions, with phrase and response alterna-
tives. Within these stories, you’re able to connect outside APIs to bring data into the chatbot. These
features are the right mix of user friendliness to allow anyone to build a bot, while also allowing for the
maximum amount of customization.
Links
Like Chatfuel, Botsify permits template linking similar to the CNN use case. These links allow informa-
tion to easily be tied to various communications between bot and user.
Converse and Learning
The Converse and Learning tools in Botisfy bring the AI capabilities of this system to a new level. In
this window, you’re able to “teach” your bot based on real conversations within the application that
it wasn’t able to answer.

43
Templates
The templates section of the Botsify application will make marketers swoon. These templates feature
various ways to maximize interaction with chatbot users.
With templates, you can allow various messages to be tied directly into product options that link to
purchase opportunities.
API.AI
API.AI is another third-party tool that permits chatbot creation; however, the focus here is not firmly
centered around Facebook Messenger as the application platform. API.AI is able to build conversa-
tional interfaces with a number of applications and devices.

44
Below, you’ll find the way API.AI layers into the chatbot development process.
As you can see, API.AI serves as a hybrid model between custom-coded chatbots and full third-party
service providers, giving you a ton of flexibility.
API.AI receives a query as either text in natural language or an event. The system matches the query
to the most suitable intent based on information contained in the information the user enters into
the “Intent” module. This area of API.AI is similar to Botsify’s “Stories.” The process of transforming
natural language into actionable data is called Natural Language Understanding (NLU).
We’ve covered the subject of creating your own chatbot in quite some detail in this chapter, and we
offered you some examples of several tools to consider when developing a chatbot on your own. In
the next chapter, we’ll move beyond chatbots a bit and get you on the path to putting machine learn-
ing to work for you in your marketing efforts.

How Marketers Can
Get Started with
Machine Learning
CHAPTER 8

46
Now that we’ve taken the time to outline exactly how you can use and build
chatbots, let’s take a look at specific third-party options that let you get
started with machine learning regardless of your technical prowess.
MonkeyLearn
MonkeyLearn, a machine learning API that uses machines to understand human language, offers a
solution to create and test machine learning algorithms to solve problems such as sentiment analysis
and topic detection. You’ll find pre-created modules in the application, but you can also create your
own customized modules.
The modules represent three components: classification, extraction, and pipelines.
Classification
These modules take text and return labels organized in a hierarchy. Let’s take a look at its pre-built
retail classifier.
The algorithm for classification was taught over time by having “samples” added to the system. To see
the classification in action, you can classify “Text” or a “File.” We entered a description of a fishing rod
and reel combo to test the functionality.
Results come back in a JSON format, but they’re still readable by people with a base level of develop-
ment understanding.
HOW MARKETERS CAN GET STARTED WITH MACHINE LEARNINGCH. 8MACHINE LEARNING FOR MARKETERS

47
The probability rate gets graded on a 0 to 1 point scale, and several categories get returned that can
share space in a hierarchy. Below is a sample of the data:
{
“category_id”: 18313026,
“probability”: 0.152,
“label”: “Fishing”
},
{
“label”: “Fishing Rods & Reels”
},
{
“label”: “Fishing Rod & Reel Combos”
As you can see, Fishing Rod & Reel Combos was the category with the highest probability.
Extraction
You can’t create a customized text extractor today, but you can train pre-built modules. The extractors
allow you to extract keywords, useful data, and entities. We’ll look below at the useful data extractor.

48
This tool allows people to easily extract valuable information from a block of text and store it. This
type of functionality could be used to help customer service representatives manage potential issues
that arise through email or social media. Post-sentiment analysis for this module could grab the infor-
mation needed and store it on a customer data level.
Pipelines
The pipeline modules allow you to combine other modules into one workflow. For example, given a
text, you can create a pipeline that performs the following functions:
• First, it does language detection.
• Second, if the language is English, then it does sentiment analysis in English.
• Third, if the sentiment is negative, it extracts valuable data.
With pipelines, you can declare the different processing steps made to the text and create logic to
combine the partial results.
The advantage of using a pipeline is that you can do multiple process with only one HTTP request.
This option is more technical in that you need to use JSON to tie together other modules you’ve
already built into a workflow process. Alternatively, you can turn to a useful tool called Zapier, a way
to make various apps connect with one another. When you allow these apps to connect, you can save
yourself valuable time by automating repetitive, tedious tasks. For example, with Zapier, you can unite
a Trigger (“a new email”) with a specific action (“create a contact”) to execute an action in one app
when something occurs in another app.
Amazon Machine Learning
Amazon launched its machine learning as a service product in 2015, and while the product isn’t as
intuitive to use as some other third-party applications, it’s a fantastic supervised learning option
for predictions. This system uses a logistic regression model, mapping variables between 1 and 0.
Additionally, Amazon can also handle multiclass classifications. Multiclass classification models can
do the following:

49
• Automate the effort of predicting object classification based on one of more than
two classes
• Create a scalable way to account for classes and objects that are continually evolving
Prepare Your Data
You’re going to work with data stored in an Amazon S3 location, so you need to become familiar
with uploading data through the Amazon Management Console. You’ll get an S3 URL once you’ve
uploaded your data; however, you should be aware that the S3 Browser is slightly more intuitive than
the Amazon Management Console.
The data set can hold as many attributes as you want. An attribute in this model works as a classifier.
Regardless of which attributes you include in your dataset, you must include the attribute y. This is the
training data for your model and will later be the attribute you’re looking to predict.
Since we’re using a logistic regression model, we want to choose questions where the answers have
only two choices (yes and no or true and false). For your training data, you can represent these two
choices with 1s and 0s. For multiclass classifications, you’ll use as many identifiers as you need to rep-
resent as many classes as you have set up for classifying your data.
Create a Training Source
Go into your machine learning console and choose the standard setup under “Get started …”
On the next page, you’ll need to enter the data set location that you entered in the previous step. Next,
you go through a number of steps where you verify the data, allow Amazon ML to infer the schema,
and select the y as the target attribute for predictions.
Create an ML Model
The default model for the ML model is the option that Amazon creates based on its understanding of
your schema and your predictive attribute. However, you can customize the recipe yourself. If you’re
getting started with machine learning, it would be best to experiment with the default models and get
an understanding of the system first before you venture into recipe creation.
Set a Score Threshold
Once you’ve created the model, Amazon will tell you if your data and recipe combined are good
enough to use. Amazon computes an Area Under a Curve (AUC) Metric that expresses the quality

50
performance of your machine learning model. The closer the AUC for a machine learning model
comes to 1, the better the quality of the performance.
Amazon lets you know in real-world language whether the data is good or bad.
You can now use the score threshold to fine-tune your ML model performance. You can use this score
to target the most true or false items predicted based on likelihood.
Generate Predictions
Now you’re ready to generate predictions.
You can generate a real-time prediction to test the model. Here, you simply enter the attributes for
one record and see what the prediction for the record returns.

51
You’re also able to do batch predictions. For this process, you’ll need to upload the data you intend
to analyze to S3 and choose Batch Predictions in the Machine Learning menu. You’ll then be asked to
enter the S3 URL, verify the data, and name your batch prediction. The process will then run for you.
To view your data, go back to Batch Predictions in the Amazon Machine Learning menu.
A list of predictions will appear, and you choose the batch prediction you ran.
You want to note the Output S3 URL. Go to the Amazon S3 console, navigate to that referenced URL’s
location, and find the results folder. The results are in a compressed .gzip file. You can download the
file, open it, and analyze the results.
The data has two columns. One is for best answer, which delivers a true or false 0 or 1 based on the
score threshold you set. The other column has the raw score of the record.

52
Application
The application for this type of service is fairly open, and that flexibility is what makes this option such
an interesting one for marketers.
One option is to glean more from your marketing campaigns. You can take the data from your ana-
lytics and create attributes for users that entered through your campaigns as they worked their way
through the funnel. As you launch new campaigns, you can measure success in early days by running
a model that compares your data to the performance of previous campaigns. You could also set up
these models for different levels of your funnel to tell your salespeople about records that they need
to focus on at each level. This arrangement is referred to as lead scoring. When applied to machine
learning, you essentially “train” a machine to recognize individuals in a marketing database that have
the potential to become future customers. The machine assigns predictive lead scores to identify the
likelihood that each individual account will convert to a sale or a customer.
Google Cloud Prediction API
Google Cloud Prediction API provides a RESTful API to build machine learning models. You can build
spam detection, sentiment analysis, and recommendation systems based on the models.
This system is similar to Amazon’s Machine Learning option, except it involves a much more advanced
knowledge of basic coding skill to get up and running. However, you’ll find an option for those of us
that have less refined experience creating machine learning recipes from syntax.
Google Cloud Prediction API has a Google Sheets option called Smart Autofill.
Just like our other processes, the first step starts with data. You pull in the data based on important
attributes for your campaign. Unlike the process we did with Amazon, you actually want to combine
your training data and data you want to predict in the same sheet.
From the data you enter, you choose the columns that have data you want to use in predicting your
results, as well as the column you plan to fill with predictions.

53
Next, you select the column that has the data you want to predict.
When you’re done, you can view the estimated error and accuracy of your data. You can work with
values that are not simply true or false statements, and the add-on function will report the possible
amount of error.
We ran the Auto Smartfill add-on on the data from the Amazon Machine Learning test. The test
worked, but we noticed a couple of items.
• As with traditional Google Sheets operations, the add-on has a hard time dealing with
data at a very large scale. We had to parse down the record count to get it to work.
• It takes some time to get the results you need.
However, the results you get are very good.

54
Microsoft Azure ML
Microsoft has launched ML Studio as a part of its Azure tool set.
ML Studio is interesting because you can work in a drag-and-drop interface to build your workflows.
While this arrangement makes the setup look extremely user-friendly, the robust nature of ML Studio
is not something anyone can dive into. Each of the interconnected units in ML Studio is a layer of the
machine learning process. This layering includes data, algorithms, scoring models, and more. Each of
those components is capable of being managed through R or Python.
As you dig in, you realize that Microsoft Azure ML Studio is a powerful tool created with data scientists
in mind. This feature makes it a good option if you’re growing a data science team, but it’s less friendly
for a new user to the system.
One interesting note is that you’ll discover pre-built solutions available in the Azure marketplace that
might make this option more palatable for those with technical experience.
Out-of-the-Box Tools
While the above options allow for an entire world of flexibility and creativity for your machine learning
needs, some of us only need tools that work today for very specific issues. Luckily, machine learn-
ing-based software solutions are on the rise in the adtech space, and solutions exist for you to sign up
today and put machine learning to work for you.
Atomic Reach
Atomic Reach is an out-of-the-box solution that helps marketers know what to do with their content
to get the most value out of their content marketing. This tool helps you create the content, share it,
engage with data in real time to make on-the-spot changes to strategy, and see trending topics that
you can use to create timely content.
The creation module measures 22 attributes of the content you’re creating in real time. These attri-
butes include readability level, length, emotions and more. The Atomic Reach Writer lets you train
your machine learning by assigning different attribute levels as an “audience.” You generate content
for that audience, and the system lets you know where you score.

55
Rare.io
Rare.io has created an email marketing optimization solution that works with Shopify and
BigCommerce. The company’s main offering is predictive email recommendations. You can create
and use preset Mailouts based on visitor segmentations:
You’re able to test different subject lines and email templates to see how different customer segmen-
tations will likely react to them based on the predictive analysis of Rare.io.
Open-Source Options
Before we close the door on the topic of tools and ways that marketers can get working with
machine learning today, we wanted to touch on some open-source platforms that exist and are worth
investigating.
Keras.io
Keras is an open-source neural network library written in Python. It’s capable of running on top of
three open-source machine learning software libraries, DeepLearning4j, TensorFlow, and Theano. The
library has many implementations of commonly used neural network building blocks including layers,
objectives, and optimizers.
Orange
Orange is an open-source data visualization, machine learning, and data-mining toolkit. It has many
uses, but its widgets are dedicated to the following:
• Data: Data input, filtering, and manipulation
• Classification: Supervised Machine Learning algorithms for classification
• Regression: Supervised Machine Learning algorithms for regression
• Evaluate: Cross-validation, sampling-based procedures, and reliability scoring
• Unsupervised: Unsupervised Learning algorithms for clustering and
data projection techniques

56
Orange is a great learner tool because it allows anyone to jump in and test out machine learning. Even
though it is great for beginners, Orange allows you to embed deep learning networks and other out-
side resources. For example you can utilize ImageNet to to image classification :
The best part of Orange is it’s easy to use drag and drop interface.
The steps to run basic linear regression are very straightforward.
Needed Signals
Inputs:
• A data set
• Preprocessor
• A preprocessed data set.
Outputs:
• Learner
• A linear regression learning algorithm with settings as specified in the dialog.
• Predictor
• A trained regressor. Output signal sent only if input Data is present.
First, you will drag and drop the Linear Regression Widget. You are able to customize the widget to
customize the learner.

57
1. You will name the predictor
2. You can choose anyone of the existing models
3. You can produce a report
4. Save your changes
Next, you will setup a workflow from your data to Linear Regression. In the below example, you can
see that you can run two models simultaneously, in this case random forest is being run in parallel.
Knime
KNIME (Konstanz Information Miner) is an open-source data analytics, reporting, and integration
platform that has components of machine learning included in its modular data pipelining. KNIME
integrates with various open-source projects to help with machine learning functionality, such as
Weka and R Project.

Most Effective
Machine Learning
Models
CHAPTER 9

59
Previously, we’ve already talked about the macro levels of machine learn-
ing (ML) which included the supervised and the nonsupervised module.
Additionally, we also discussed the best tools which marketers use for ML
and the best way of using such tools. In this article, we’ll be talking about
the models and the algorithms which stand as inner logic for all those that
we’ve previously discussed.
Having a thorough knowledge on the models that build up the core of each tool will give you more
options when using those tools.
Regression Models
Linear Regression
The primary task of linear relationship models is to establish the relationship between two different
factors or variables. The variable or factor being predicted is the dependent variable; the factors used
to predict values of the dependent variable would be the independent variables.
Establishing the relationships between dependent and independent variables is represented by the
equation ya*x+b where:
y = the dependent variable
a = the slope
x = the independent variable
b = the intercept
The intercept and the slope are determined by minimizing the total sum of the square difference of
the regression line and the data points.
Simplifying this formula, we can say: Linear regressions predict values through the existing variables
and their relationship with one another.
A Sample Case in Marketing Use
A value which this particular model could have is predicting the value of possible client personas
being targeted toward a new product release. As soon as you have broken down your client personas
— including their buying patterns — you will be able to pull out data points which represent corre-
sponding client personas to substitute in the linear regression equation. This arrangement will provide
a precise prediction for how much a client persona may try to purchase.
The moment marketers use a linear regression as a way to estimate sales out of each persona, they
can also estimate the marketing channel value to offer a better estimate on the personas for each
channel. This optimization will then optimize the marketing campaign even before its launch.
Logistic Regression
Unlike actual regression, this type of regression isn’t focused on predicting the numeric variable
value through a set of inputs. Alternatively, it delivers a probability that the provided input point is
coming from a specific class. In logistic regression, we are simply looking for probabilities that an
event will happen.
MOST EFFECTIVE MACHINE LEARNING MODELSCH. 9MACHINE LEARNING FOR MARKETERS

60
Logistic regression focuses on the assumption that the input space could be set apart to two “regions”
through a linear boundary.
The two region answers would be similar to
True/False
Yes/No
The tools that use logistic regression within their core include Google’s prediction API along with an
ML application from Amazon.
With the case mentioned earlier, we could use a logistic regression to determine if a sale has a likeli-
hood to close, and we are able to use the linear regression to determine the amount of money which
clients are likely to spend. Using this information, we can use a lead list which follows:
Linear: decides how much money a lead will be likely to spend
Logistic: determines the chances of clients spending that money
Next come up with a forecast for your overall lead list using that combination of amount and
probability.
Model Clustering
K-Means Clustering Algorithm
K-means is among the popular unsupervised ML algorithms designed for clustering analysis. It is a non-
deterministic and iterative method. Its algorithm works over a specified data set using a predefined
set of clusters, k. The result of a k-means algorithm is the k clusters of input data partitioned from
the clusters.
As an example, let’s have a k-Means clustering application for Wikipedia search results. We’ll use the
search term Lincoln over Wikipedia. When we do, we will get results containing this word, and they
usually give us references to the past U.S. president, a city in the United States, or a car company. We
can use this algorithm to put these results and pages into groups which talk about the same concepts.
Thus, the algorithm will group the results and pages into clusters.
A k-means clustering algorithm is commonly used by search engines to identify similarities of search
results. This algorithm provides lesser time for the searcher when looking for a more precise result
using the search engine.
K-means clustering stands out to be an effective tool when developing marketing and buying personas
out of clusters of information concerning buyers.
Cluster analysis proves to be useful for market segmentation. This analysis means having to divide a
market’s potential customers over to different subsets wherein customers that belong to the same
group are the same concerning a specified characteristic and vice versa. This arrangement will provide
a convenient and more accurate marketing mix depending on the target customer specified.

Thorough
knowledge on
the models that
build up the
core of each tool
will give you
more options.
MOST EFFECTIVE MACHINE LEARNING MODELS

62
Model Classifications
Decision Trees
Decision trees are graphical representations which use branching methodology to deliver all possible
results of a certain decision using a few specified conditions. Within a single decision tree, the partic-
ular internal node gets represented as a test over an attribute. Each corresponding branch of the tree
will represent the possible outcome of the test. The leaf node represents the particular class label
such as the decision made after all attributes have been compute. Classification rules will therefore be
represented via the path from the root toward the leaf node.
The types of decision trees include the following:
• Classification trees: The default type of decision trees is used to set apart data sets into
two varying classes using a response variable. This arrangement is generally used when
a categorical nature is detected out of a response variable.
• Regression trees: This type of decision tree is used if the result or target variable will be
numerical or continuous — mostly used for predictive type of issues.
Two classifications of decision trees exist, depending on the type of the target variable. These two
types are composed of the binary variable decision tree and the continuous variable decision tree.
Such ML algorithms allow for reliable decisions in uncertainties and help improve communication
by providing a visual representation for a decision situation. Decision-tree ML algorithms provide
assistance to data scientists by capturing the idea that when one decision gets implemented, its oper-
ational nature will be something specific.
Decision-tree algorithms give developers the best decisions by letting them traverse forward and
backward in a calculation path.
We can use decision trees referring to the various cases we’ve mentioned previously.
Regarding classification, trees gets used in classification of data sets through response variables. Using
this model, we are able to use the actions of historical data from previous website visitors to classify
potential customers over to various lead funnel buckets.
Regression trees will then give a prediction for whether a client will make a purchase using the
actions provided.
Random Forests
This ML algorithm uses the bagging approach to create a number of decision trees using random sub-
sets of data. One model gets trained a number of times over random samples of data sets to obtain
high accuracy prediction over the random forest algorithm. Within this method, all outputs combine
to come up with the last final decision. This final prediction gets computed by creating a poll of the
results from each decision tree.
Random forests are much like decision trees. We can use them to predict client behaviors. The only
difference is that random forests use a number of decision trees and give marketers a higher level
of accuracy.

63
Naïve Bayes Classifier Algorithm
Classifying lengthy text notes is almost close to impossible when done manually. A Naïve Bayes classi-
fier algorithm comes into place for this occasion. A classifier serves as a function which assigns a value
for a population using the categories available. An example of this algorithm is the spam filtering in
emails where the filter becomes a classifier that assigns an email as either spam or not spam.
This classifier is among those common learning methods grouped by similarities which makes use
of Bayes’ theorem of probability to build ML models, especially those related to disease prediction
and document classification. This is merely a basic classification of words that is on account of Bayes
probability theorem with regards to subjective analysis of any content.
Facebook uses sentiment analysis to determine status updates showing a negative or positive emotion.
Google uses document categorization to index documents and find relevancy scores such as PageRank.
A Naïve Bayes algorithm can be used for article news classification for politics, sports, technology,
and other topics.
Google also uses email spam filtering to decide whether to classify emails as spam or through the
Naïve Bayes algorithm.
Support Vector Machine Algorithm (SVM)
This supervised ML algorithm design is mainly for classification and regression issues in which a certain
data set instructs the SVM about the classes so that it can classify fresh data sets. This classification is
achieved by classifying the assigned data sets to different classes through linear hyperplanes. The SVM
algorithm will try to maximize the certain distance of the different classes involved, known as margin
maximization. If the line that maximizes such distance will be identified, the possibility to generalize
well over unseen datasets will be increased.
Two types of SVMs
Linear SVMs are the training data which include the classifiers separated by the hyperplane.
Nonlinear SVMs concern the classifiers’ inability to become separated by the hyperplane. An exam-
ple of this situation is face detection which comprise a number of images composed of faces and
non-faces (images other than faces). Through specific conditions, the training data becomes too com-
plicated, becoming impossible to trace a representation for each feature vector. Having to separate
the set of faces linearly through the set of non-faces is going to be a complex task.
Social Media SVMs get used to track clear mentions of a certain brand. The SVMs classify them from
among other brands. Some of the metrics include loyalty, likes, and perceived value.
Through theme generation through social media, the algorithm can map social mentions over to
predefined categories specific to an industry. An example of this mapping is the hospitality and
travel labels which point out key themes discussed over social media regarding a company’s ser-
vices or products.
Email routing SVMs automatically classify emails over clients directed to complaint handlers. This
classification reduces turnaround times and offers quick solution to issues.

64
Opinion mining is the computational learning of the attitude of people, their opinions, and their
stances on issues, companies, individuals, or services. SVMs are used to classify the polarity among
texts in a document or sentence level in the classifications: negative, positive or neutral.
Artificial Neural Networks (ANN)
The artificial neuron network (ANN) serves as a computational model based over a structure and
function of a biological neural network. Data which flows in the network will affect the structure of an
ANN since the neural network varies — or becomes smarter in a sense — out of the output or input.
ANNS are nonlinear statistic data modeling platforms where the intricate relationships among out-
puts and inputs become modeled or where patterns exist.
Additionally, ANN is a neural network. The ANN comes with a few advantages; one of them is learning
just by observing data sets. Using ANN as a random function approximation tool, you can estimate the
cost effectiveness and best methods to arrive over a solution during the process of defining comput-
ing functions and distributions.
ANNs come with three interconnected layers. The initial layer comes with input neurons where such
neurons deliver data over to the second layer and then over to a third layer.
Having to train an artificial neural network will require choosing over an allowed model in a pool of
associated algorithms.
Ad targeting uses a neural network to efficiently decide where to deploy advertising spending. One
example is Baidu which handles ad targeting. It launched the use of artificial neural networks.
This Chinese web giant also uses deep learning to target ads over its many lines of services. As Andrew
Ng, a co-launcher of the deep learning operation of Google and Baidu’s chief research officer, “It’s
used very successfully in advertising.” He further adds, “We have not released revenue numbers on
the specific impact, but it is significant.”
Your model selection will be decided by the data you are working with, the tools your team utilizes,
and the questions you are trying to answer. As you can see, an ML model is not an one size fits all
tool. Luckily for marketers the decision of what model fits with which activity is being made by ML
and AI tool developers that are making it easier for professionals across occupations to deliver
results using data science.

How to Deploy
Models Online
CHAPTER 10

66
Marketers and businesses who want to use machine learning (ML) beyond
the tools previously discussed may require better customization to deploy
their specific online models. In addition, marketers as well as business orga-
nizations may be able to come up with ways on how their developers can
fully use their models which, in this case, may call for other helpful tools.
Self-Managed and Hosted Tools
Cloudera Oryx
Oryx is primarily designed to work on Apache Hadoop. For those who are not familiar with Hadoop,
this open-source software platform is designed for storing data sets and executing applications over
groups of commodity hardware. It can store a substantial amount of data of any kind, provides a high
level of processing power, and is capable of handling a virtually limitless amount of tasks or jobs.
Oryx 2 is a product of Lambda Architecture built upon Apache Kafka and Apache Spark; however, Oryx
2 is designed with a specialty in real-time machine learning at a large scale. This framework allows for
easy creation of applications, but also comes with packaged, end-to-end apps designed for collabora-
tive filtering, regression, clustering, and classification.
From Oryx.io
Oryx is composed of three tiers: Each tier builds over the one below:
1. The generic Lambda Architecture tier provides serving, batch, and speed layers that
are not particular to ML.
2. Specialization over top provides ML simplifications for hyperparameter selecting and
other selections.
3. The end-to-end execution of similar standard machine learning algorithms gives the
application (k-means, random decision forests, or Alternating Least Squares Method
for Collaborative Filtering) over the top.
HOW TO DEPLOY MODELS ONLINECH. 10MACHINE LEARNING FOR MARKETERS

67
Oryx Implementation
Execute the three layers using:
./oryx-run.sh batch
./oryx-run.sh speed
./oryx-run.sh serving
You don’t have to use these layers over one machine; however, the application can also be possible if
the configuration specifies various ports to accommodate the speed and batch layer Spark web user
interface along with the port for serving layer API. You can run the serving layer over a number of
machines.
Say the batch layer Spark user interface runs on the 4040 port of the machine where you started it —
that is, unless this arrangement was altered by a configuration. As a default, the port 8080 will come
with a web-based console designed for the serving layer.
Example of Use
A sample GroupLens 100K data set is found in a u.data file. The data set should be converted to CSV:
Provide the input over a serving layer, having a certain local command line tool such as a curl:
In case you are actually tailing the particular input topic, you will find substantial CSV data flow toward
the topic:
After a few moments, you will find the batch layer triggering a new computation. This sample config-
uration will start per five-minute intervals.
The data provided is first delivered to HDFS. The sample configuration has it delivered to the directory
hdfs:///user/example/Oryx/data/. Located inside are directories named by a time stamp with each one
having a Hadoop part — files equipped with input being SequenceFiles of a text. Despite not being pure
text, if you have to print them, you should be able to deliver recognizable data since it is actual text.

68
Afterward, a model computation will take place. This computation should show results in the form of
a number of fresh distributed jobs over the batch layer. The Spark UI will be staired over http://your-
batch-layer:4040 as provided in the sample configuration.
The model will then be complete and will be considered as a blend of PMML and various supporting
data in the subdirectory of hdfs:///user/example/Oryx/model/. Here’s an example: The model PMML
files will be PMML files that contain elements such as the following:
The Y/ and X/ subdirectories that go with it come with feature vectors such as:
If you’re tailing the particular update topic, you will find these values to be published over the topic.
This publication will then be detected by the serving layer later and will provide a return status at the
/ready endpoint of 200 OK:

69
Apache Mahout
Three primary features of Apache Mahout include the following:
• A basic and extensible coding environment and framework designed for building scal-
able algorithms
• A wide array of premade algorithms for H2O, Apache Flink, and Scala + Apache Spark
• Samsara, the vector mathematics experimentation setting having R-like syntax that
works at scale
Considering that Mahout’s primary algorithms are designed for classification, you can implement clus-
tering as well as batch-based collaborative filtering over Apache Hadoop through the map/reduce
paradigm. However, this process does not prevent any contributions over to Hadoop-based imple-
mentations. The contributions which operate over one node or over a non-Hadoop cluster become
adapted or welcomed. As an example, the “taste” collaborative-filtering influencer part of Mahout was
originally a single project and runs stand-alone without the aid of Hadoop.
An algebraic back end-independent optimizer — along with an algebraic Scala DSL to unify in-memory
and distributed algebraic operators — makes up the setting. During this time, the available algebraic
platforms supported include H2O, Apache Flink, and Apache Spark. MapReduce algorithms have grad-
ually lost support from users.
H2O
H2O is an open-source, scalable ML and a predictive analytics system which provides convenience in
building ML models over substantial data. H2O also offers effortless productionalization over models
used in an enterprise.
This platform’s code is written in Java language. Within H2O, a certain DistributedKey/Value store gets
used for approach as well as reference models, objects, data, and other elements over all the machines
and nodes.
The algorithms are implemented over H2O’s dispersed Map/Reduce framework. After implementa-
tion, you make use of the Java Fork/Join framework to deliver multi-threading. The particular data gets
read in parallel and gets distributed all over the cluster and stored inside a memory within compressed
columnar format. Additionally, the H2O data parser comes with a built-in intelligence that guesses the
schema of a certain incoming data set and offers data ingest support over a number of sources over
various formats.
The REST API from H2O provides access to every capability of H2O out of both external program and
script out of JSON on HTTP. The H2O web interface, R binding, and Python binding uses REST API.
Accord.NET Framework
For developers as well as marketers who are presently in the .net working environment, Accord
Framework stands out as a robust open-source alternative when deploying models.
Accord.NET Framework is actually a .NET ML framework combined with image and audio-processing
libraries coded in C#. Accord.NET is a comprehensive framework used to create production-grade
computer vision, signal processing, statistical applications, and computer audition.
MLlib
MLlib is one popular ML library consisting of basic learning algorithms as well as utilities that include
regression, clustering, collaborative filtering, dimensionality reduction, and classification.

70
This platform conveniently adapts to Spark’s APIs and works conveniently with NumPy from Python
and R libraries. Hadoop data sourcing can also be used on this platform, a function which allows it to
work effectively with Hadoop workflows as well.
Fully Managed Tools
Yhat
Yhat is a Y Combinator-supported enterprise delivering an end-to-end data science platform for
deploying, managing, and developing real-time decision APIs.
This platform conveniently takes away the pains brought about by IT obstacles and problems which
come with cloud-based data science such as server setting and configuration. Through this platform,
data scientists and experts are able to convert static insights into available decision-making APIs which
can work effortlessly with any client- and employee-facing application. This platform also brought about
the birth of Rodeo, an open-source IDE (integrated development environment) for the Python platform.
Developers have also come up with the product ScienceOps, a tool which aims to provide data sci-
ence users, experts, and developers the capability to seamlessly and quickly deploy models in any
environment. You use these models in cases involving lead scoring, image recognition, credit analysis,
and various other predictive models.
Developers, data scientists, and experts use Python or R to easily and quickly deploy models through
Yhat instance. Coders and developers are then able to make use of a REST API in embedding models
over to web applications. This embedding allows for real-time ML application. One excellent use case
of such technology surrounds the submission and approval of real-time credit applications.
Below is an example of a model deployed through Yhat.
To use Yhat, you generate your model within Python or R session, change the user name and API key
fields, and then run the script.

71
The Yhat platform comes with a basic Beer Recommender which provides a simple use case of the
platform to generate accurate predictions based on the specific data that a user enters.
Now that we’ve looked at how marketers with an advanced understanding of ML can develop and
deploy models online, it’s time to consider how data scientists take modeling to the next level in the
next chapter of this guide.

How Data Scientists
Take Modeling to
the Next Level
CHAPTER 11

73
Data scientists are exceptional people who bring data analysis to a new level.
These analytical experts have a high level of technical skills and can solve
the most complicated problems with a high level of inquisitiveness.
In simpler words, data scientists are those who structure and deploy the specific models which we’ve
been discussing. In the previous discussions, we’ve talked about platforms, and those platforms are
where data scientists deploy their specific models. However, they can also use other tools to analyze
their data.
Roles of a Data Scientist
• Pinpoint which data-analytics problem would offer the most benefit to an organization.
• Identify the ideal variables and data sets.
• Gather multiple sets of unstructured and structured data out of disparate sources.
• Verify and clean data for increased accuracy, uniformity, and completeness.
• Come up with algorithms and models to hunt the stores of large data.
• Analyze data for pattern and trend identification.
• Interpret data to trace ideal solutions and opportunities.
• Share findings with stakeholders through the use of visualization models
and other methods.
From research conducted by Burtch Works about data scientists, 88 percent of them have a master’s
degree while 46 percent have doctoral degrees. Not surprisingly, most data scientists are highly edu-
cated people.
Languages
Given that the primary role of data scientists is to connect the world of IT to the world of mathe-
matics, they need to understand how to deploy their code and how to use technology to deliver
their analyses. The most common coding languages which data analysts use today are R Code and
Python. Other alternatives include Scala and Java; however, R and Python have become the norm in
this profession.
R vs. Python for Data Science
Both R and Python have practical applications for data scientists. Below you’ll find some highlights of
the comparisons between R and Python and when you may use each one.
R is useful when the data analysis you plan to do requires you to dig in for some preliminary data
dives, and it’s useful for almost any type of data analysis you’d want to do because of the large number
of packages and readily accessible tests that can provide you with the tools you need to get up and
moving on your data analysis. R packages (see below) can further your reach with packages for data
visualizations and machine learning.
Python is useful when you need to integrate your data analysis activities with web apps or if you need
to incorporate some statistics codes into a production database arrangement. As a robust program-
ming language on its own, Python makes for a practical way to put algorithms into place when you’re
working with production applications.
In the past, Python packages for data analysis were in their infancy, but their growth and capabili-
ties have increased over recent years. NumPy, SciPy and pandas for data manipulation (see more
HOW DATA SCIENTISTS TAKE MODELING TO THE NEXT LEVELCH. 11MACHINE LEARNING FOR MARKETERS

74
in Python libraries below) have made Python useful from a data analysis perspective. From an ML
standpoint, scikit-learn is useful when you want to consider multiple versions of a single algorithm.
Why Do Data Scientists Recommend Python?
Among those high-level languages commonly used by data scientists, Python makes the claim that it’s
easy to learn, friendly, fast, open, and powerful. Developed in the 1980s and named after the comedy
group Monty Python, this platform has established its presence in the data science industry because
of its number of libraries created using this language.
Additionally, Python also has a simple syntax which is easy to understand for coding newbies, one
that is easily recognizable for anyone who has dealt with Java, Visual Basic, C/C++, and Matlab. Python
prides itself on being a multipurpose, user-friendly programming language when dealing with quanti-
tative and analytical computing.
Data scientists also find Python an easy language to use when developing visualizations that serve as
an essential function to their work while relaying analysis to organizational stakeholders. This process
gets carried out using APIs such as data visualization libraries and Plotly.
Another selling point of Python is that it scales quickly and gets equipped with handy tools such as
IPython Notebook – known today as Jupyter Notebook. This tool serves as an interactive computa-
tional setting where rich text, plots, rich media, mathematics, and code execution can be joined. You
can also run a number of blocks and lines of code over various cells using this tool. Additionally, you
can play with them, interchange their positions up or down, and have results display just below the
cell. Lastly, writing R, Scale, SQL, and other languages in this tool is also possible which permits easier
and more efficient workflow.
Why Do Data Scientists Recommend R Code?
R is another open-source coding language and platform setting developed for graphics and statistical
computing. It is R Foundation for Statistical Computing supported.
Developed in the 1990s by Ross Ihaka along with Robert Gentleman from New Zealand’s University of
Auckland, R served as statistical software for students. It was further improved through the decades
through the ever-increasing number of user-created libraries.
R Code can be used in many libraries and data scientists prefer it because of these other reasons:
• R serves as an object-oriented platform developed by statisticians. R creates objects,
functions, and operators, all of which give users a convenient way to model, visualize,
and explore data.
• Using R, standard statistical methods become easy to implement. Since most predictive
and statistical modeling today is already done in R, most techniques get introduced
first using R.
• R is free and comes with a high level of numerical accuracy and quality thanks to count-
less improvement efforts by developers through the years. Having the R open interface
offers easy system and application integration.

75
The R environment provides the following:
• Efficient storage facility and data handling
• A selection of operators to make calculations on specific arrays
within particular matrices
• A manageable set of tools used for data analysis
• Graphical models designed for data analysis
These reasons can support why R is an excellent choice among data scientists over other platforms
such as Python. That is, R can offer a better ability to deploy, visualize, and develop models within
one system.
Packages and Libraries
Packages and libraries give data scientists the tools inside Python for faster scaling of visualizations
and data analysis.
Python Libraries
In the Python Tutorial, Modules are defined as files composed of Python statements and definitions
— the filename equipped with the suffix .py.
Packages offer a way to structure the Python’s module namespace by “dotted module names.”
“Library” is a term used for a number of codes applied for use in various applications. It delivers
generic functionality for specific applications.
The moment a package or module gets published, those entities can be called by many as a library.
For much of the time, a library can be composed of packages or a single package, but it can still be
considered as a single module.
Among those Python libraries commonly used by data scientists are the following:
NumPy
NumPy is a foundational library used for particular computing within Python. This platform intro-
duces models for multidimensional matrices and arrays. Additionally, routines enable developers to
execute advanced statistical and mathematical functions using as little coding as possible.
SciPy
This platform builds upon NumPy through accumulating algorithms along with high-level commands
designed for visualizing and manipulating data. It also comes with functions designed for solving dif-
ferential equations, computing integrals numerically, and many other functions.
Pandas
Pandas offer tools and data structures for data analysis in statistics, finance, social sciences, and engi-
neering. This platform is convenient to use over unlabeled, messy, and incomplete data, and it allows
shaping, merging, slicing, and reshaping of data sets.
IPython
This platform extends the functionality of the interactive interpreter of Python. IPython accomplishes
this functionality through a souped-up interactive shell combining a shell syntax, rich media, and tab
completion, plus a command history retrieval. Additionally, this platform also serves as an embeddable
interpreter for various programs used for debugging.

76
Matplotlib
This platform is essential in a Python library when you’re developing 2D plots and graphs. Compared
to the more advanced libraries, Matplotlib needs more commands to create better-looking graphs
and figures.
Scikit-learn
This platform can build over SciPy and NumPy by combining algorithms for common machine learn-
ing (ML) and data mining tasks, along with clustering, classification, and regression. It is a curated
library that permits easy selection of various versions of a single algorithm.
Theano
This library uses a NumPy-like syntax to optimize and evaluate mathematical expressions. It comes
with an amazing speed that is highly essential for deep learning and complex computational tasks.
TensorFlow
Developed by Google, this platform is a high-profile entrant in the field of ML. TensorFlow became
developed as a successor to the open-source platform DistBelief, a framework for modeling and train-
ing neural works.
Scrapy
This aptly named platform is designed for developing spider bots which crawl over the web and
extract structured data such as contact info, prices, and URLs.
NLTK
This platform compiles several libraries designed for NLP, Natural Language Processing. NLTK (Natural
Language Toolkit) enables easy entity identification, text tagging, and parse trees display. In addition,
you can also do more complex tasks such as automatic summarization and sentiment analysis.
Pattern
This platform combines NLTK and Scrapy functionality and acts as an out-of-the-box solution for web
mining, ML, NLP, and network analysis. It comes with tools such as APIs for Google, Wikipedia, Twitter,
and text-analysis algorithms.
Seaborn
Seabornisanothercommonvisualizationlibraryusedbymanydatascientists.ItbuildsoverMatplotlib’s
foundation and is an easier tool to use for generating various types of plots along the lines of violin
plots, time series, and heat maps.
Bokeh
This tool permits zoomable and interactive plots over modern web browsers through JavaScript wid-
gets. It also provides a three-level interface over high-level abstractions which enable you to generate
complex plots quickly.
Basemap
This tool allows you to easily manipulate and transfer maps over to Matplotlib by taking Matplotlib’s
coordinates and using them over 25 different projections.
NetworkX
NetworkX permits easy creation and analysis of graphs and networks. This platform works conve-
niently on standard and non-standard data formats.

Data scientists
can solve the
most complicated
problems with
a high level of
inquisitiveness.
HOW DATA SCIENTISTS TAKE MODELING TO THE NEXT LEVEL

78
R Packages
R Packages are a collection of R functions, compiled code, and data stored over a well-defined format.
As soon as you install this tool, load it over to a session so that you can use it.
forecast
If you need to consider time series, or a series of observations about well-defined data points obtained
through repeated measurements over time, you’ll find the forecast package useful for time series
analysis. This package can also help you when you need to predict the future value of a time series
(for example, a future stock price).
nnet
This common platform is widely used because it is easy to understand. Unfortunately, nnet can also be
a challenge to use since the platform gets limited only to one layer of nodes.
klaR Visualization and Classification
The CARET (Classification and Regression Training) module combines prediction and model train-
ing. This combination allows data scientists to run multiple algorithms for a specific business
problem. Additionally, developers can investigate parameters for a specific algorithm through con-
trolled experiments.
plyr
This package offering is like the salvation of a cool breeze on a sultry summer day. When you need to
break up a big data structure into homogenous units, apply a function to each one of those pieces, and
then bring all the results back together, plyr has the capability to fit the same model to each subset
of a data frame, calculate summary statistics for each group, and then perform transformations that
include scaling and standardizing.
ggplot2
Answering the pleas of those who work with graphics, ggplot2 offers a powerful model of graphics
creation that allow you to shape complex multi-layered graphics. This package operates on the idea
that you can build graphs by using a data set, visual marks (called gemos) that represent data points,
and a coordinate system. It also takes care of those details like creating legends for graphics.
igraph
This platform comes equipped with a set of tools used for network analysis.
Random Forest
As one of the widely used algorithms in the ML field, this platform can be used to develop multiple
decision trees where you’ll input observations. When using this platform, data scientists must use
numeric variables or factors. You’re allowed a maximum of 32 levels when using Random Forest.
reshape2
Shifting data between long and wide formats gets less complicated with reshape2. Using two func-
tions, melt and cast, you can take wide-format data and render it as long-form data (melting) or take
long-format data and present it as wide-format data (casting). With reshape2, the format your data
takes follows from what you intend to do with your data.
We’ve covered a plethora of the tools behind ML in this chapter. While these tools are sophisticated
ones for marketers, we certainly want you to be aware that not everything in life is 100 percent fool-
proof. In the next chapter, we’ll drill down into some cautionary tales about ML problems and how you
can improve your success and accuracy with ML.

Common
Problems with
Machine Learning
CHAPTER 12

80
Machine learning (ML) can provide a great deal of advantages for any mar-
keter as long as marketers use the technology efficiently. Knowing the
possible issues and problems companies face can help you avoid the same
mistakes and better use ML.
Common Practical Mistakes
Focusing Too Much on Algorithms and Theories
Leave advanced mathematics to the experts. As you embark on a journey with ML, you’ll be drawn in
to the concepts that build the foundation of science, but you may still be on the other end of results
that you won’t be able to achieve after learning everything. Fortunately, the experts have already taken
care of the more complicated tasks and algorithmic and theoretical challenges. With this help, master-
ing all the foundational theories along with statistics of an ML project won’t be necessary.
Mastering ALL of ML
Once you become an expert in ML, you become a data scientist. For those who are not data scientists,
you don’t need to master everything about ML. All you have to do is to identify the issues which you
will be solving and find the best model resources to help you solve those issues.
For example, for those dealing with basic predictive modeling, you wouldn’t need the expertise of a
master on natural language processing. Why would you spend time being an expert in the field when
you can just master the niches of ML to solve specific problems?
Using Changing or Premade Tools
Previously, we’ve discussed the best tools such as R Code and Python which data scientists use for
making customizable solutions for their projects. For the nonexperts, tools such as Orange and
Amazon S3 could already suffice. With these simple but handy tools, we are able to get busy, get work-
ing, and get answers quickly. All that is left to do when using these tools is to focus on making analyses.
When you have found that ideal tool to help you solve your problem, don’t switch tools. Many devel-
opers switch tools as soon as they find new ones in the market. Although trying out other tools may
be essential to find your ideal option, you should stick to one tool as soon as you find it. Don’t play with
other tools as this practice can make you lose track of solving your problem.
Having Algorithms Become Obsolete as Soon as Data Grows
ML algorithms will always require much data when being trained. Often, these ML algorithms will be
trained over a particular data set and then used to predict future data, a process which you can’t
easily anticipate. The previously “accurate” model over a data set may no longer be as accurate as it
once was when the set of data changes. For a system that changes slowly, the accuracy may still not
be compromised; however, if the system changes rapidly, the ML algorithm will have a lesser accuracy
rate given that the past data no longer applies.
Getting Bad Predictions to Come Together With Biases
ML algorithms can pinpoint the specific biases which can cause problems for a business. An example
of this problem can occur when a car insurance company tries to predict which client has a high rate
of getting into a car accident and tries to strip out the gender preference given that the law does not
allow such discrimination. Even without gender as a part of the data set, the algorithm can still deter-
mine the gender through correlates and eventually use gender as a predictor form.
COMMON PROBLEMS WITH MACHINE LEARNINGCH. 12MACHINE LEARNING FOR MARKETERS

81
With this example, we can draw out two principles. The first you need to impose additional constraints
over an algorithm other than accuracy alone. Second, the smarter the algorithm becomes, the more
difficulty you’ll have controlling it.
Developers always use ML to develop predictors. Such predictors include improving search results
and product selections and anticipating the behavior of customers. One reason behind inaccurate
predictions may be overfitting, which occurs when the ML algorithm adapts to the noise in its data
instead of uncovering the basic signal.
When you want to fit complex models to a small amount of data, you can always do so. Doing so will
then allow your complex model to hit every data point, including the random fluctuations. Depending
on the amount of data and noise, you can fit a complex model that matches these requirements.
Marketers should always keep these items in mind when dealing with data sets. Make sure that your
data is as clean of an inherent bias as possible and overfitting resulting from noise in the data set.
You can deal with this concern immediately during the evaluation stage of an ML project while you’re
looking at the variations between training and test data.
Making the Wrong Assumptions
ML algorithms running over fully automated systems have to be able to deal with missing data
points. One popular approach to this issue is using mean value as a replacement for the missing
value. This application will provide reliable assumptions about data including the particular data
missing at random.
Whether they’re being used in automated systems or not, ML algorithms automatically assume that
the data is random and representative. However, having random data in a company is not common.
The best way to deal with this issue is to make sure that your data does not come with gaping holes
and can deliver a substantial amount of assumptions.
Receiving Bad Recommendations
Recommendation engines are already common today. While some may be reliable, others may not
seem to be more accurate. ML algorithms impose what these recommendation engines learn. One
example can be seen when a customer’s taste changes; the recommendations will already become
useless. Experts call this phenomenon “exploitation versus exploration” trade-off. In the event the
algorithm tries to exploit what it learned devoid of exploration, it will reinforce the data that it has, will
not try to entertain new data, and will become unusable.
To deal with this issue, marketers need to add the varying changes in tastes over time-sensitive
niches such as fashion. With this step, you can avoid recommending winter coats to your clients
during the summer.
Having Bad Data Convert to Bad Results
Not all data will be relevant and valuable. If data is not well understood, ML results could also provide
negative expectations. The initial testing would say that you are right about everything, but when
launched, your model becomes disastrous. When creating products, data scientists should initiate
tests using unforeseen variables, which include smart attackers, so that they can know about any
possible outcome.
Have your ML project start and end with high-quality data. Having garbage within the system automat-
ically converts to garbage over the end of the system.

Machine learning
can provide a great
deal of advantages
for any marketer as
long as marketers
use the technology
efficiently.
COMMON PROBLEMS WITH MACHINE LEARNING

83
Machine Learning Goes Wrong
Despite the many success stories with ML, we can also find the failures. While machines are constantly
evolving, events can also show us that ML is not as reliable in achieving intelligence which far exceeds
that of humans.
Below are a few examples of when ML goes wrong.
Microsoft and the Chatbot Tay
Microsoft set up the chatbot Tay to simulate the image of a teenage girl over Twitter, show the
world its most advanced technology, and connect with modern users. The company included what it
assumed to be an impenetrable layer of ML and then ran the program over a certain search engine to
get responses from its audiences. The developers gave Tay an adolescent personality along with some
common one-liners before presenting the program to the online world. Unfortunately, the program
didn’t perform well with the internet crowd, bashed with racist comments, anti-Semitic ideas, and
obscene words from audiences. In the end, Microsoft had shut down the experiment and apologized
for the offensive and hurtful tweets.
With this example, it would seem that ML-powered programs are still not as advanced and intelligent
as we expect them to be. Research shows that only two tweets were more than enough to bring Tay
down and brand it as anti-Semitic. However, in Tay’s defense, the words she used were only those
taught to her and those from conversations in the internet. In light of this observation, the appropri-
ateness filter was not present in Tay’s system.
Uber
Uber has also dealt with the same problem when ML did not work well with them. This ride-sharing
app comes with an algorithm which automatically responds to increased demands by increasing its
fare rates. During the Martin Place siege over Sydney, the prices quadrupled, leaving criticisms from
most of its customers. The app algorithm detected a sudden spike in the demand and alternatively
increased its price to draw more drivers to that particular place with high demand. ML understood
the demand; however, it could not interpret why the particular increased demand happened. With this
example, we can safely say that algorithms need to have a few inputs which allow them to connect to
real-world scenarios.
These examples should not discourage a marketer from using ML tools to lessen their workloads.
These tales are merely cautionary, common mistakes which marketers should keep in mind when
developing ML algorithms and projects. Machine learning models are constantly evolving and the
insufficiency can be overcomed with exponentially growing real-world data and computation power
in the near future.

Machine Learning
Quick Start
CHAPTER 13

85
A certain notion exists that machine learning is only for the experts and is
not for people with less knowledge of programming. The people you’ll be
dealing with are machines, and machines can’t make decisions on their own.
Your job is to make these machines think. At first, you would think that this
task is impossible, but through this guide, you’ll understand how machine
learning (ML) can be much easier than it sounds.
As with almost every framework in the history of humankind — along the lines of Newton’s laws of
motion, the law of supply and demand, and others — the particular ideas and concepts of machine
learning are likewise achievable and not complicated. Among those challenges that you can face when
discovering machine learning are the notations, formulas, and a specific language which you may have
never heard about before.
Choosing Your Data Set
Just about any data set can be used for ML purposes, but an efficient model can only be achieved
through a well-structured data set.
When thinking about data sets, the better data sets with well-labeled data that features a number of
ML algorithms mean you won’t have a hard time building your model. However, coming up with sim-
pler algorithms is best despite the volume, especially when you’re dealing with a number of data sets,
for easier sorting later on. As an example, regularized logistic regression using billions of features can
work well despite the substantial numbers.
The more complicated algorithms — which include GBRT (Gradient Boosted Regression Tree), deep
neural networks, and random forests — perform better than simpler platforms; however, they can be
more expensive to use and train. When you’re working with tons of data, you will always make use of
linear perceptron and other learning algorithms online. When using any learning algorithms online,
you always need to use a stochastic gradient descent.
Perhaps always choosing a good selection of features would be safe, such as the essential tools to
learn about your data. Also a features selection serves more likely an art instead of a science. The
selection process could be complicated; however, it could be eased out if you work on removing the
unpromising ones first to narrow down your options.
Model Selection
Tofindthebestmodelthatsuitsyourneeds,youneedtokeepinmindafewofthefollowingconsiderations:
Accuracy
While accuracy is important, you don’t always have to go with the most accurate answer. This can
happen when you feel that you have a model that is producing a high bias or high variance. Sometimes
an approximation will do depending on the use. Opting for the approximation, you’ll be able to reduce
processing time and you avoid overfitting.
Training Time
Each algorithm will require varying training times for instructing its models. As we consider the train-
ing time, we should note that this time is closely related to the accuracy — as training time increases,
so does the accuracy. Also keep in mind that other algorithms can be more sensitive than others and
may require more training time.
MACHINE LEARNING QUICK STARTCH. 13MACHINE LEARNING FOR MARKETERS

86
Linear Classification
Linearity is common in machine learning algorithms. A line can possibly separate classes as assumed
by linear classification algorithms. These classes can be composed of support vector and logistic
regression machines. Linear classification algorithms work off of assumptions which can lead to the
previously mentioned high bias and variance issues. These algorithms are still commonly implemented
first because they’re algorithmically simple and quick to train.
Number of Parameters
Data scientists use precise measurements to set up an algorithm called parameters. These parameters
are in the form of numbers which provide changes on the behavior of algorithms. These behavioral
changes can include the number of iterations, error tolerances, or options among variants about
the behavior of algorithms. Sometimes, if you want to have the best settings, you may often need to
adjust the accuracy and the training time. The algorithms having the largest number of parameters will
always require more trial-and-error combinations and tests.
How You Can Train Your Model
A substantial amount of repetition will always be required with training a model to achieve a high level
of accuracy out of a selected model.
As a first step, you’ll have to select the best model suited for your particular project. Previously, we
discussed the pros and cons for various models. With the information provided, the steps would cover
the following:
Data Set Partitioning for Testing and Training
Also called the process of “data partitioning,” you’ll find various options from the variety of tools and
languages to help you choose which data points to divide over training and testing data sets. Below is
an example of partitioning data sets over R.
1 # By default R will come with a number of datasets
2 data = mtcars
3 dim(data) # 32 11
4
5 #Sample Indexes
6 indexes = sample(1:nrow(data), size=0.2*nrow(data))
7
8 # Split data
9 test = data[indexes,]
10 dim(test) # 6 11
11 train = data[-indexes,]
12 dim(train) # 26 11

87
Eliminating Biases in Data Through Cross-Validation
Performing a cross-validation out of your model will give you a guarantee for an accurately correct
model that doesn’t pick too much noise or, in simpler terms, low on variance and bias. Cross-validation
becomes a reliable technique if you want to track the effectiveness of your particular model especially
within cases where you need to ease overfitting.
You will find various cross-validation techniques available where the holdout method could be the
most basic form that makes use of the partitioning method. Function approximators will only fit a
certain function through training sets and afterward will be prompted to make a prediction on the
output values for the particular data within the testing set. One good feature about this method is its
quickness to compute, a benefit recommended for the residual method.
Maintain the Accuracy of Your Model
Using the data provided after a cross-validation process, let’s check out the errors and devise a way
to fix these errors.
Repeat These Steps Until You Obtain the Desired Accuracy
Gather and Test
When on the lookout for the basic data sets to learn with and gain comfort with, experts would
recommend multivariate data sets provided by Ronald Fisher, a British statistician who, in 1936, intro-
duced the Iris Flower Data Set. This data set studied four measured features across three different
species of iris. It comes with 50 samples for every species.
Moving on to the next steps as soon as we acquire our data set:
Basic Predictive Model Out of Orange
Let’s go with a pair of basic predictive models within the Orange canvas along with their specific
Jupyter notebooks.
Initially, let’s examine our Simple Predictive Model – Part 1 notebook. After we do, let’s proceed with
creating the model over the Orange Canvas. Below is a diagram for predicting results coming from
the Iris data set by using a classification from within Orange.
The toolbar over the left part of the canvas provides you with more than 100 widgets that you can use
by dragging the toolbar over the canvas. Reading the schema from left to right, and details from widget

88
to widget out of the pipelines, the Iris data set can be seen over a variety of widgets once it is loaded in.
By clicking the simple data table along with the scatter plot, we’ll have the following displayed:

89
Using three widgets, we can already have a better picture of our data. The scatter plot provides us
with the option “Rank Projections,” which tracks the best step to view the specific data. In this regard,
having “Petal Width vs. Petal Length” provides a quick pattern on the width of the flower’s petal and
the type of Iris flower studied.
Looking at the way we build the predictive model, we connect the data over a classification tree
widget to examine through the viewer widget.
Here the predictive model becomes noticeable and thus allows us to connect the model plus data
over to the “Test and Score” along with the “Predictions” widgets. This process illustrates how the
classification tree performs.
The Predictions widget will be able to predict the specific type of iris flower through the set of input
data. Using a confusion matrix, we can find the most accurate predictions.

90
Hold Out Test Over Orange to Build Predictive Model
In this next example for predictive modeling, we’ll be having a slightly complicated model out of hold-
ing out a particular test set. Through this effort, we can use varying data sets to test and train the
model, thus disabling overfitting.
Using Orange Canvas, we can now proceed with building the same predictive model for a better view
of what we are building.
Are you with us so far? Now let’s take your knowledge to the next level and apply what you’ve learned
in this chapter. We’ll next consider some more resources you can use to get you started with a more
complex ML project.

Machine learning for Marketers

More Related Content

What's hot (6)

Similar to Machine learning for Marketers (20)

Recently uploaded (20)

Machine learning for Marketers