MeasureFest July 2021 - Session Segmentation with Machine Learning

Audience segmentation with
machine learning
Richard Lawrence
Rise at Seven
@richlawre

About me
SEO background,
studying a Data
Science degree in
spare time.
@richlawre

About me
Follow me on Twitter
@richlawre
@richlawre

What we’re going to
cover
@richlawre

@richlawre
A bit of context
about machine
learning AUDIENCE 1
Pageviews
Transaction
revenue
AUDIENCE 2
AUDIENCE 3
The agenda

@richlawre
An overview of
how audience
segmentation
works
AUDIENCE 1
Pageviews
Transaction
revenue
AUDIENCE 2
AUDIENCE 3
The agenda

@richlawre
Some detail about
how to do it
AUDIENCE 1
Pageviews
Transaction
revenue
AUDIENCE 2
AUDIENCE 3
The agenda

@richlawre
How to take things
further
AUDIENCE 1
Pageviews
Transaction
revenue
AUDIENCE 2
AUDIENCE 3
The agenda

It learns with
labelled data
@richlawre
What is supervised
machine learning?

@richlawre
It finds
patterns in
data
What is unsupervised
machine learning?

Audience segmentation
in a nutshell
@richlawre

We extract data about
individual sessions from
web analytics
@richlawre

Extracting the data
CHANNEL SESSIONS TRANSACTIONS REVENUE
Organic search 1000 50 £12,000
Paid search 700 30 £3,000
Direct 500 25 £6,000
Referral 300 30 £4,000
Instead of
grouping sessions
by channel or
section...
@richlawre

Extracting the data
...we extract
details about
individual
sessions
@richlawre
SESSION ID PAGEVIEWS TIME PER PAGE REVENUE
Session 1 7 30 seconds £77.50

We then use unsupervised
machine learning to find
interesting patterns
@richlawre

Instead of
analysing sessions
grouped together
in some way...
Finding patterns
@richlawre

AUDIENCE 1
Pageviews
Transaction
revenue
AUDIENCE 2
AUDIENCE 3
...we use machine
learning to find
patterns in user
behaviour.
@richlawre
Finding patterns

This results in
actionable audience
segments
@richlawre

The Gatherer
Landing section: Homepage
Least time
per page
Most number of
pages viewed
Highest number of
conversions per
session
Most likely to
download a
brochure
Description:
The Gatherer comes directly to the website to the homepage,
visits multiple car models to download a brochure for each to look
at offline later.
Example CRO Test:
Link to a model comparison table from
the homepage with option to download
a brochure for each model
Likely onsite journey
Example segment from
Car manufacturer
Second section: Car Models
Exit section: Car Models
@richlawre

The Skipper
Example segment from
Train operator
Description:
The Skipper has likely already done their travel research (around
when to travel & where) multiple times without buying and are
simply returning - likely at the last minute - to finally finish task.
Example CRO Test:
Use a cookie to add a banner to the
homepage that takes a returning user
back to where they left off in the
transaction process.
Slightly more time
per page then average
More likely to buy in
the evening or at night
Fewest days since last
session
Fewest pages per
visit
Over index for
visiting via tablet
Over index for
visiting via email
@richlawre

Why do you need to do
this?
@richlawre

1.Find behaviours you
may not have realised
existed
@richlawre

2.Generate test
hypotheses for CRO
@richlawre

3.Track behaviours
over time (more about
this later)
@richlawre

The key steps
1. Extract the data
2. Process the data
3. Select features
4. Cluster the data
5. Manually explore the segments
@richlawre

1.Extracting the data
@richlawre

Using Google Analytics API
Extract by Session ID
or Client ID
@richlawre
https://www.jcchouinard.com/google-analytics-api-using-python/

Useful dimensions:
landingPagePath
secondPagePath
exitPagePath
@richlawre

Useful metrics:
pageviewsPerSession
revenuePerTransaction
goalXXCompletions
@richlawre

There is a limit on the
number of
metrics/dimensions
10
@richlawre

There is also a limit
on the number of
rows per call
25,000
@richlawre

The answer is to
loop over days,
metrics, dimensions
& merge!
@richlawre

Using BigQuery
Data is nested -
I’ve found it makes
things more
difficult at the
session level
https://adswerve.com/blog/google-analytics-queries-in-bigquery-part-two-users-sessions-unnesting-hits/
@richlawre

Using BigQuery
However it is
possible to do and
there is some great
information around
@richlawre

Using BigQuery
Can also run the
unsupervised
machine learning
algorithm directly
in SQL
@richlawre

Using BigQuery
Previously used 1M
sessions with
Python & Google
Colab - BigQuery
wasn’t necessary
@richlawre

Using BigQuery
Choose days at
random to ensure
variation
@richlawre

2.Processing the data
@richlawre

Useful data transformations
Change hours of
the day to
morning,
afternoon,
evening,night
SESSION ID DAY DAY TYPE
Session 1 Monday Weekday
Session 2 Tuesday Weekday
Session 3 Saturday Weekend
Session 4 Wednesday Weekday
@richlawre

Change days to
weekday &
weekend
@richlawre

Change pages to
sections
@richlawre

Combine certain
conversion points
@richlawre

Here is a useful link
to do find and
replace it in Python
& Pandas
@richlawre

You could use
Google DataPrep
instead
@richlawre

One hot encoding
Converts categories
to 1s & 0s.
SESSION
ID
CHANNEL
Session 1 Organic Search
Session 2 Paid Search
Session 3 Direct
Session 4 Direct
SESSION
ID
ORGANIC
SEARCH
PAID
SEARCH
DIRECT
Session 1 1 0 0
Session 2 0 1 0
Session 3 0 0 1
Session 4 0 0 1
@richlawre

Values aren’t
increasing so doesn’t
skew the clustering
algorithm
SESSION
ID
CHANNEL
Session 3 Direct
Session 4 Direct
SESSION
ID
ORGANIC
SEARCH
PAID
SEARCH
DIRECT
Session 1 1 0 0
Session 2 0 1 0
Session 3 0 0 1
Session 4 0 0 1
@richlawre
One hot encoding

Use for numerical as
well as categorical
data
SESSION
ID
CHANNEL
Session 3 Direct
Session 4 Direct
SESSION
ID
ORGANIC
SEARCH
PAID
SEARCH
DIRECT
Session 1 1 0 0
Session 2 0 1 0
Session 3 0 0 1
Session 4 0 0 1
@richlawre
One hot encoding

See here for how to
do it with Python
SESSION
ID
CHANNEL
Session 3 Direct
Session 4 Direct
SESSION
ID
ORGANIC
SEARCH
PAID
SEARCH
DIRECT
Session 1 1 0 0
Session 2 0 1 0
Session 3 0 0 1
Session 4 0 0 1
@richlawre
One hot encoding

3.Selecting features
@richlawre

Best subset regression
Choose desired
response variable
& find potential
explanatory
variables
@richlawre

Runs regression
analysis for
combinations of
variables at once to
find correlation
@richlawre

This will help you
narrow down
features to find
useful patterns
within
@richlawre

See Python
walkthrough here
@richlawre

4.Clustering the data
@richlawre

Principal Component
Analysis
Transforms a large
set of variables into
a smaller one
without much loss
@richlawre

Principal Component
Analysis
See walkthrough
here.
@richlawre

Using a KMeans algorithm
The
unsupervised
machine learning
algorithm to find
patterns
@richlawre

See a full
walkthrough here
with Python.
@richlawre

You can also do this
directly in BigQuery.
@richlawre

Using a silhouette score
Way of finding the
optimum
number of
clusters
@richlawre

Using a silhouette score
Optimal number is
at the elbow in
the graph - not
much gain after
this
@richlawre

5.Always manually
explore the segments!
@richlawre

Taking it to the next level
@richlawre

Classify any future session
@richlawre
Use the labelled
data to train a
supervised
machine learning
algorithm - we use
deep learning

@richlawre
The better defined
your segments,
the better this will
perform

@richlawre
Push the labelled
sessions back into
Google Analytics
via Data Import

Visualise in Streamlit
@richlawre
CRM segment 1
CRM segment 2
CRM segment 3
CRM segment 4
CRM segment 5

Unsupervised machine learning
finds interesting patterns in data.
@richlawre

Apply this to individual sessions
from Google Analytics to create
behaviour segments.
@richlawre

This can be a great source of ideas
for CRO hypotheses.
@richlawre

There are 5 steps for the analysis:
extract,process,feature selection,
cluster,manually explore
@richlawre

You can use Python or other
toolsets (Google Cloud) to do the
analysis.
@richlawre

You can use the segments to label
any future session on the website.
@richlawre

MeasureFest July 2021 - Session Segmentation with Machine Learning

More Related Content

What's hot (20)

Similar to MeasureFest July 2021 - Session Segmentation with Machine Learning (20)

Recently uploaded (20)

MeasureFest July 2021 - Session Segmentation with Machine Learning