Social Media Mining - Chapter 5 (Data Mining Essentials)

Data Mining
Essentials
SOCIAL
MEDIA
MINING

2Social Media Mining Measures and Metrics 2Social Media Mining Data Mining Essentialshttp://socialmediamining.info/
Dear instructors/users of these slides:
Please feel free to include these slides in your own
material, or modify them as you see fit. If you decide
to incorporate these slides into your presentations,
please include the following note:
R. Zafarani, M. A. Abbasi, and H. Liu, Social Media Mining:
An Introduction, Cambridge University Press, 2014.
Free book and slides at http://socialmediamining.info/
or include a link to the website:
http://socialmediamining.info/

Introduction
Data production rate has increased
dramatically (Big Data) and we are
able store much more data
– E.g., purchase data, social media
data, cellphone data
Businesses and customers need
useful or actionable knowledge to
gain insight from raw data for
various purposes
– It’s not just searching data or
databases
The process of extracting useful patterns from raw data is
known as Knowledge Discovery in Databases (KDD)

KDD Process

Collecting Data on Social Media
• Collect Raw Data
– Use site provided APIs
• Flickr’s: https://www.flickr.com/services/api/
– Scrape information directly
• Use Provided Repositories
– http://socialcomputing.asu.edu
– http://snap.Stanford.edu
– https://github.com/caesar0301/awesome-public-
datasets

Data Mining
• Extracting/“mining” knowledge from large-scale data (big data)
• Data-driven discovery and modeling of hidden patterns in big data
• Extracting information/knowledge from data that is
– implicit,
– previously unknown,
– unexpected, and
– potentially useful
Data Mining: the process of discovering hidden and
actionable patterns from data
It utilizes methods at the intersection of artificial intelligence,
machine learning, statistics, and database systems

Data Mining vs. Databases
• Data mining is the process of extracting
hidden and actionable patterns from data
• Database systems store and manage data
– Queries return part of stored data
– Queries do not extract hidden patterns
• Examples of querying databases
– Find all employees with income more than $250K
– Find top spending customers in last month
– Find all students from engineering college with GPA
more than average

Examples of Data Mining Applications
• Fraud/Spam Detections: Identifying fraudulent
transactions of a credit card or spam emails
– You are given a user’s purchase history and a new transaction,
identify whether the transaction is fraud or not;
– Determine whether a given email is spam or not
• Frequent Patterns: Extracting purchase patterns
from existing records
– beer ⇒ dippers (80%)
• Forecasting: Forecasting future sales and needs
according to some given samples
• Finding Like-Minded Individuals: Extracting groups
of like-minded people in a given network

Data

Data Instances
• In the KDD process,
– Data is in a tabular format (a set of instances)
• Each instance is a collection of properties and features
related to an object or person
– A patient’s medical record
– A user’s profile
– A gene’s information
• Instances are also called points, data points, or
observations
Data Instance:
Features ( Attributes or measurements) Class Label
Feature Value Class Attribute

Data Instances
• Predicting whether an individual who visits an online
book seller is going to buy a specific book
• Features can be
– Continues: values are numeric values
• Money spent: $25
– Discrete: can take a number of values
• Money spent: {high, normal, low}
Labeled
Example
Unlabeled
Example

Data Types + Permissible Operations (statistics)
• Nominal
– Operations:
• Mode (most common feature value), Equality Comparison
– E.g., {male, female}
• Ordinal
– Feature values have an intrinsic order to them, but the difference is not
defined
– Operations:
• same as nominal, feature value rank
– E.g., {Low, medium, high}
• Interval
– Operations:
• Addition and subtractions are allowed whereas divisions and multiplications are
not
– E.g., 3:08 PM, calendar dates
• Ratio
– Operations:
• divisions and multiplications are allowed
– E.g., Height, weight, money quantities

Sample Dataset – Twitter Users
Activity Date Joined Number of Followers Verified Account?
Has Profile
Picture?
High 2015 50 FALSE no
High 2013 300 TRUE no
Average 2011 860000 FALSE yes
Low 2012 96 FALSE yes
High 2008 8,000 FALSE yes
Average 2009 5 TRUE no
Very High 2010 650,000 TRUE yes
Low 2010 95 FALSE no
Average 2011 70 FALSE yes
Very High 2013 80,000 FALSE yes
Low 2014 70 TRUE yes
Average 2013 900 TRUE yes
High 2011 7500 FALSE yes
Low 2010 910 TRUE no
NominalOrdinal Interval Ratio Nominal

Text Representation
• The most common way to represent
documents is to transform them into vectors
– Process them with linear algebraic operations
• This representation is called “Bag of Words”
– Vector Space Model
• Weights for words can be assigned by TF-IDF

Vector Space Model
• Consider a set of documents 𝐷
• Each document is a set of words
• Goal: convert these documents to vectors
How to set 𝒘𝒋, 𝒊
• Set 𝑤𝑗, 𝑖 to 1 when the word 𝑗 exists in document 𝑖 and 0
when it does not.
• We can also set 𝑤𝑗, 𝑖 to the number, of times the word 𝑗
is observed in document 𝑖 (frequency)
• 𝑑𝑖 : document 𝑖
• 𝑤𝑗, 𝑖 : the weight for word 𝑗 in document 𝑖

Vector Space Model: An Example
• Documents:
– 𝑑1: social media mining
– 𝑑2: social media data
– 𝑑3: financial market data
• Reference vector (Dictionary):
– (social, media, mining, data, financial, market)
• Vector representation:

TF-IDF (Term Frequency-Inverse Document Frequency)
TF-IDF of term (word) 𝑡, document 𝑑, and
document corpus 𝐷 is calculated as follows:
is the frequency of word 𝑗 in document 𝑖
The total number of
documents in the corpus
The number of documents where the term 𝑗 appears

TF-IDF: An Example
Document 𝑑1 contains 100 words
– Word “apple” appears 10 times in 𝑑1
– Word “orange” appears 20 times in 𝑑1
We have |𝐷| = 20 documents
– Word “apple” only appears in document 𝑑1
– Word “orange” appears in all 20 documents

TF-IDF: An Example
• Documents:
– 𝑑1: social media mining
– 𝑑2: social media data
– 𝑑3: financial market data
• TF values:
• TF-IDF

Data Quality
When making data ready for mining, data quality needs to be assured
• Noise
– Noise is the distortion of the data
• Outliers
– Outliers are data points that are considerably different from other data points in
the dataset
• Missing Values
– Missing feature values in data instances
– Solution:
• Remove instances that have missing values
• Estimate missing values, and
• Ignore missing values when running data mining algorithm
• Duplicate data

Data Preprocessing
• Aggregation
– It is performed when multiple features need to be combined into a single one or
when the scale of the features change
– Example: image width , image height -> image area (width x height)
• Discretization
– From continues values to discrete values
– Example: money spent -> {low, normal, high}
• Feature Selection
– Choose relevant features
• Feature Extraction
– Creating new features from original features
– Often, more complicated than aggregation
• Sampling
– Random Sampling
– Sampling with or without replacement
– Stratified Sampling: useful when having class imbalance
– Social Network Sampling

Data Preprocessing
Sampling social networks:
• Start with a small set of nodes
(seed nodes)
• Sample
– (a) the connected components
they belong to;
– (b) the set of nodes (and edges)
connected to them directly; or
– (c) the set of nodes and edges
that are within n-hop distance
from them.

Data Mining Algorithms

Data Mining Algorithms
• Supervised Learning Algorithm
– Classification (class attribute is discrete)
• Assign data into predefined classes
– Spam Detection
– Regression (class attribute takes real values)
• Predict a real value for a given data instance
– Predict the price for a given house
• Unsupervised Learning Algorithm
– Group similar items together into some clusters
• Detect communities in a given social network

Supervised Learning

Supervised Learning: The Process
• We are given a set of labeled records/instances
– In the format (𝑿, 𝑦)
– 𝑿 is a vector of features
– 𝑦 is the class attribute (commonly a scalar)
• [Training] The supervised learning task is to build model that
maps 𝑿 to 𝑦 (find a mapping 𝑚 such that 𝑚(𝑿) = 𝑦)
• [Testing] Given an unlabeled instance (𝑿’, ? ), we compute 𝑚(𝑿’)
– E.g., spam/non-spam prediction

Classification: An Email Example
• A set of emails is given
– Users have manually labeled
them as spam / non-spam
• Use a set of features (x) to
identify spam/non-spam
status of the email (y)
– We can use words in the email
• In this case, classes are
𝑦 = {𝑠𝑝𝑎𝑚, 𝑛𝑜𝑛 − 𝑠𝑝𝑎𝑚}

A Twitter Example

Supervised Learning Algorithms
• Classification
– Decision tree learning
– Naive Bayes Classifier
– 𝑘-nearest neighbor classifier
– Classification with Network information
• Regression
– Linear Regression
– Logistic Regression

Decision Tree Learning

Decision Tree
• A decision tree is
learned from the
dataset
– (training data with known
classes)
• The learned tree is later
applied to predict the
class attribute value of
new data
– (test data with unknown
classes)
– Only the feature values
are known

Decision Tree: Example
Class Labels
Multiple
decision trees
can be learned
from the same
dataset

Decision Tree Construction
• Decision trees are constructed recursively
– A top-down greedy approach in which features are sequentially
selected.
• After selecting a feature for each node,
– based on its attribute values, different branches are created.
• The training set is then partitioned into subsets based on
the feature values,
– each of which fall under the respective feature value branch;
– The process is continued for these subsets and other nodes
• When selecting features, we prefer features that partition
the set of instances into subsets that are more pure.
• A pure subset has instances that all have the same class
attribute value.

Decision Tree Construction
• When reaching pure (or highly pure) subsets under a branch,
– the decision tree construction process no longer partitions the
subset,
– creates a leaf under the branch, and
– assigns the class attribute value (or the majority class attribute value)
for subset instances as the leaf’s predicted class attribute value
• To measure purity we can use/minimize entropy.
• Over a subset of training instances, 𝑇, with a binary class
attribute (values in {+,-}), the entropy of 𝑇 is defined as:
– 𝑝+ is the proportion of positive examples in 𝑇
– 𝑝− is the proportion of negative examples in 𝑇

Entropy Example
Assume there is a subset 𝑇 that has 10 instances:
• Seven instances have a positive class attribute value
• Three have a negative class attribute value
• Denote 𝑇 as [7+, 3-]
• The entropy for subset 𝑇 is
In a pure subset, all instances have the same class attribute
value (entropy is 0)
If the subset contains an unequal number of positive and
negative instances
• The entropy is between 0 and 1.

Naïve Bayes Learning

Naive Bayes Classifier
For two random variables 𝑿 and 𝑌, Bayes
theorem states that,
class variable the instance features
Then class attribute value for instance 𝑿
We assume that features
are independent given the
class attribute

NBC: An Example

Nearest Neighbor Classifier

Nearest Neighbor Classifier
• 𝑘-nearest neighbor or 𝑘-NN,
– Utilizes the neighbors of an instance to perform classification.
• It uses the 𝑘 nearest instances, called neighbors, to
perform classification.
• The instance being classified is assigned the label (class
attribute value) that the majority of its 𝑘 neighbors are
assigned
• When 𝑘 = 1, the closest neighbor’s label is used as the
predicted label for the instance being classified
• To determine the neighbors of an instance, we need to
measure its distance to all other instances based on some
distance metric.
– Often Euclidean distance is employed

𝑘-NN: Algorithm

𝑘-NN example
• When 𝑘 = 5, the predicted label is:
• When 𝑘 = 9, the predicted label is:
∆
□

𝑘-NN: Example 2
Data instance Outlook Temperature Humidity Similarity Label K Prediction
2 1 1 1 3 N 1 N
1 1 0 1 2 N 2 N
4 0 1 1 2 Y 3 N
3 0 0 1 1 Y 4 ?
5 1 0 0 1 Y 5 Y
6 0 0 0 0 N 6 ?
7 0 0 0 0 Y 7 Y
Similarity between row 8 and other data instances;
(Similarity = 1 if attributes have the same value, otherwise similarity = 0)

Classification with
Network Information

Classification with Network Information
• Consider a friendship network on social media and a
product being marketed to this network.
• The product seller wants to know who the potential buyers
are for this product.
• Assume we are given the network with the list of
individuals that decided to buy or not buy the product. Our
goal is to predict the decision for the undecided
individuals.
• This problem can be formulated as a classification problem
based on features gathered from individuals.
• However, in this case, we have additional friendship
information that may be helpful in building better
classification models

Classification with Network Information
• Let 𝑦𝑖 denote the label for node 𝑖.
• We can assume that
• How can we estimate P(𝑦𝑖 = 1|𝑁 𝑣𝑖 )?

Weighted-vote Relational-Neighbor (wvRN)
wvRN provides an approach to estimate P(𝑦𝑖 = 1|𝑁 𝑣𝑖 )
• In wvRN, to find the label of a node, we compute a
weighted vote among its neighbors
• P(𝑦𝑖 = 1|𝑁 𝑣𝑖 ) is only calculated for unlabeled 𝑣𝑖’s
• We need to compute these probabilities using some
order until convergence [i.e., they don’t change]
– What happens for different orders?

wvRN example

Regression

Regression
In regression,
• Class values are real
numbers as class values
– In classification, class
values are categorical
𝑦 ≈ 𝑓(𝑋)
Features
(regressors)
𝑋 = (𝑥1, 𝑥2, … , 𝑥 𝑚)
Class attribute
(dependent variable)
𝑦𝑅
Goal: find the relation between 𝑦
and vector 𝑋 = (𝑥1, 𝑥2, … , 𝑥 𝑚)

Linear Regression
• Linear regression: we assume the relation
between the class attribute 𝑌 and feature set 𝑿 is
linear
• 𝑊 represents the vector of regression coefficients
• Regression can be solved by estimating 𝑊 and
 using the provided dataset and the labels 𝑌
– “Least squares” is a popular method to solve regression
– The goal is to minimize

Least Squares
Find 𝑊 such that minimizing ǁ𝑌 − 𝑋𝑊ǁ2 for
regressors 𝑿 and labels 𝑌

Simplifying W with SVD
SVD of 𝑿

Logistic Regression
• A probabilistic view of regression
• Assuming the class attribute is binary, logistic
regression finds the probability 𝑝
• We can assume that 𝑝 can be obtained from 𝑋
• Unbounded: 𝑋 can take any values and 𝛽 is
also unconstrained
𝑋: Feature vector

Logistic Regression
• Solution: transform 𝑝 using 𝑔(𝑝), such that
𝑔(𝑝) is unbounded
– Fit 𝑔(𝑝) using 𝛽X
• Known an as the logit function
• For any 𝑝 between [0,1]
– 𝐺(𝑝) is in range [−∞, +∞]

Logistic Regression
Logistic Function
– Acts as a probability
• Goal: Find 𝛽 such that 𝑃(𝑌|𝑋) is maximized
– No closed form solution
– Iterative maximum likelihood
• Prediction: once 𝛽 is determined compute 𝑃(𝑌|𝑋)
– For a binary class problem if it is more than 0.5, predict 1

Supervised Learning
Evaluation

Evaluating Supervised Learning
• Training/Testing Framework:
– A training dataset (i.e., the labels are known) is used to train a model
– the model is evaluated on a test dataset.
• The correct labels of the test dataset are unknown,
– In practice, the training set is divided into two parts,
• One used for training and
• The other used for testing.
• When testing, the labels from this test set are removed.
– After these labels are predicted using the model, the predicted
labels are compared with the masked labels (ground truth).

• Dividing the training set into train/test sets
– Leave-one-out training
• Divide the training set into 𝑘 equally sized partitions
– Often called folds
• Use all folds but one to train and the one left out for testing
– 𝑘-fold cross validation training
• Divide the training set into 𝑘 equally sized sets
• Run the algorithm 𝑘 times
• In round 𝑖, we use all folds but fold 𝑖 for training and fold 𝑖 for
testing.
• The average performance of the algorithm over 𝑘 rounds
measures the performance of the algorithm.

• As the class labels are discrete, we can measure the
accuracy by dividing number of correctly predicted
labels (𝐶) by the total number of instances (𝑁)
• More sophisticated approaches of evaluation
• AUC
• F-Measure

Evaluating Regression Performance
• The labels cannot be predicted precisely
• We need to set a margin to accept or reject
the predictions
– Example. When the observed temperature is 71,
any prediction in the range of 71±0.5 can be
considered as a correct prediction
• Or, we can use correlation between predicted
labels and the ground truth.

Unsupervised Learning

• Clustering is a form of
unsupervised
learning
• Clustering algorithms
group together
similar items
• The algorithm does not
have examples showing
how the samples should
be grouped together
(unlabeled data)
Unsupervised division of instances
into groups of similar objects
Kernel Density Estimation

Measuring Distance/Similarity in Clustering
• Clustering Goal: Group together similar items
• Instances are put into different clusters based
on the distance to other instances
• Any clustering algorithm requires a distance
measure
The most popular
(dis)similarity measure for
continuous features are
Euclidean Distance and
Pearson Linear Correlation
Euclidean Distance

Similarity Measures
• 𝑋 and 𝑌 are 𝑛-dimensional vectors
Once a distance measure is selected, instances are grouped using it.

Clustering
• Clusters are usually represented by compact and
abstract notations.
• “Cluster centroids” are one common example of
this abstract notation.
• Partitional Algorithms (most common type)
– Partition the dataset into a set of clusters
– Each instance is assigned to a cluster exactly once
– No instance remains unassigned to clusters.
• Example: 𝑘-means

𝑘-means for 𝑘 = 6

𝑘-means
The most commonly used clustering algorithm
– Related to Expectation Maximization (EM) in statistics.

𝑘-means: Algorithm
Given data points 𝑥𝑖 and an initial set
of 𝑘 centroids 𝑚1
1
, 𝑚2
1
,…, 𝑚 𝑘
1
the algorithm proceeds
as follows:
• Assignment step: Assign each data point to the
cluster 𝑆𝑖
𝑡
with the closest centroid
– Each data point goes into exactly one cluster
• Update step: Calculate the new means to be
the centroid of the data points in the cluster
– After all points are assigned

When do we stop?
The procedure is repeated until convergence
• Convergence:
– Whether centroids are no longer changing
– Equivalent to clustering assignments not changing
– The algorithm can be stopped when the Euclidean
distance between the centroids in two consecutive
steps is less than some small positive value

𝑘-means (alternative!)
• As an alternative, 𝑘-means implementations try to
minimize an objective function.
– Example: the squared distance error:
– 𝑥𝑗
𝑖
is the 𝑗th instance of cluster 𝑖
– 𝑛(𝑖) is the number of instances in cluster 𝑖
– 𝑐𝑖 is the centroid of cluster 𝑖.
• Stopping Criterion:
– when the difference between the objective function
values of two consecutive iterations of the 𝑘-means
algorithm is less than some small value .

𝑘-means
• Finding the global optimum of the 𝑘 partitions is
computationally expensive (NP-hard).
• This is equivalent to finding the optimal centroids
that minimize the objective function
• Solution: efficient heuristics
– Outcome: converge quickly to a local optimum that
might not be global
– Example: running 𝑘-means multiple times
• Select the clustering assignment that is observed most often or
• Select the clustering that is more desirable based on an
objective function, such as the squared error.

Evaluation

Evaluating the Clusterings
• Evaluation with ground truth
• Evaluation without ground truth
We are given two
types of objects
• In perfect
clustering,
objects of the
same type are
clustered
together.

Evaluation with Ground Truth
When ground truth is available,
– We have prior knowledge on what the clustering
should be (the correct clustering assignments)
• We will discuss these methods in community
analysis chapter

Evaluation without Ground Truth
• Cohesiveness
– In clustering, we are interested in clusters that
exhibit cohesiveness
– In cohesive clusters, instances inside the clusters
are close to each other
• Separateness
– We are also interested in clusterings of the data
that generates clusters that are well separated
from one another

Cohesiveness
• Cohesiveness
– In statistics: having a small standard deviation, i.e.,
being close to the mean value
– In clustering: being close to the centroid of the cluster

Separateness
• Separateness
– In statistics: separateness can be measured by
standard deviation
• Standard deviation is maximized when instances are far
from the mean
– In clustering: cluster centroids being far from the
mean of the entire dataset

Silhouette Index
• We are interested in clusters that are both cohesive and
separate
– Silhouette index
• It compares
– the average distance value between instances in the same cluster
To
– the average distance value between instances in different clusters
• In a well-clustered dataset,
– the average distance between instances in the same cluster is
small (cohesiveness) and
– the average distance between instances in different clusters is
large (separateness).

Silhouette Index
• For any instance 𝑥 that is a member of cluster 𝐶
• Compute the within-cluster average distance
• Compute the average distance between 𝑥 and
instances in cluster 𝐺
– 𝐺 is closest to 𝑥 in terms of the average distance
between 𝑥 and members of 𝐺

Silhouette Index
• Our interest: clusterings where 𝑎(𝑥) < 𝑏(𝑥)
• Silhouette can take values between [−1,1]
• The best case happens when for all 𝑥,
– 𝑎(𝑥) = 0, 𝑏(𝑥) > 𝑎(𝑥)

Silhouette Index - Example

Social Media Mining - Chapter 5 (Data Mining Essentials)

More Related Content

What's hot (20)

Similar to Social Media Mining - Chapter 5 (Data Mining Essentials) (20)

Recently uploaded (20)

Social Media Mining - Chapter 5 (Data Mining Essentials)

Editor's Notes