The document describes how a company uses various techniques to analyze user and product data from online retailers:
1. They monitor metrics and user behavior to model customers and measure site performance.
2. Predictive modeling is used to predict customer mood based on early clickstream data and recommend garments using graph models of user swapping behavior.
3. Deep learning extracts visual features of products from images to map similar garments and detect fashion trends based on user try-on preferences. Dimensionality reduction is applied to the feature space for recommendations.
6. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
7. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
11. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail
user experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment
recommendation
12. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail
user experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
16. 16
User and Garment Flow through Retailer Website
New
Users
Garments
Website
User Flow
Garment Flow
New Popular Discount ...
Unsold
Garments
Dream
No Revisit
Users
Sold
Garments
Returned
Garments
Revisit
Users
Browse Buy
…
(Try?)
17.
18.
19.
20. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
21. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
25. 25
A browser or a buyer?
Problem: 4 out of 5 customers are
new to the site
How do you decide what user
experience they need in the first 90
seconds?
Predicting Customer Mood Instagram, Facebook “You look fabulous!”
Glamorous, trendy
“I love this”
Details
Price
Fit
26. 26
Predict customer mood using Metail clickstream data
S SALE
S
Conversion
Non-Conversion
We know some information at first click...
...and a lot of information by the end of the
session...
...but can we make an
accurate prediction just
based on the first 5 events?
27. 27
Identified 174 features of
importance
Machine Learning approach to detect customer mood
Turning the
model around
to view tried
on clothes
Likely to convert
Unlikely to convert
Facebook
marketing
source
Tested predictive accuracy
Random
chance
Accuracy
of early detection
model is nearly
as good as full
session data
28. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Aggregating and describing user data
4) Predictive modelling
a) Predicting customer mood (logistic
regression)
b) Graph models and garment recommendation
29. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Aggregating and describing user data
4) Predictive modelling
a) Predicting customer mood (logistic regression)
b) Graph models and garment recommendation
31. 31
Look for events where a user removes one garment and
replaces it with another
Aggregate count across all users
Calculate net flow into each garment e.g. if 10 users
swap garment A for B, 15 swap B for A then net we
have 5 users swapping from B to A
Ignore swaps with small number of garments
Result: a simple collaborative filter
‘Other users who tried this garment also were interested
in this other garment’
Example at right shows this for Tops on a sample retailer
Garment Recommendation from User Behaviour
32. 32
Looking at multiple garment
categories, can we see any
trends based on try on
preferences?
On the right we see some
indication of users moving
away from the most visible
garment at top of page and
towards red summer dresses
We can see the garment
features visually from the
thumbnails but can we
automatically detect them?
Fashion Trend Detection
33. 33
Deep Learning: Extract Garment Features from Images
SKU image SKU features SKU feature vectors
~100 dimension
DNN Unstack
34. 34
Check whether the Garment
Attribute Predictor features are
reasonable
It looks like visually similar garments
are close together in t-SNE
Garment recommendation:
For any given SKU find ‘close’ SKUs
Outfit recommendation:
For any given SKU find ‘close’ SKUs
that are of a different garment
category
Problem: feature space has high
dimension so ‘close’ is tricky
Solution: dimensionality reduction
Dimension reduction of feature space
35. 35
Style Space Decomposition
A
B
½
User SKU
Project onto SKUs
Probability of SKUs being related maps to distance in Style
Space
Use optimization to find mapping from ~100 dimensional feature
space to ~10 dimensional Style Space
See 'Image-based recommendation on styles and substitutes' by Julian McAuley et al.
36. How we do it:
1) Monitoring and Measuring with Metrics
2) User Modelling to improve retailer and Metail user
experience
3) Predictive modelling
a) Predicting customer mood (logistic regression)
b) Graph models and garment recommendation