4. Understanding Products
To make relevant recommendations
we need to understand the products
and how they fit together
Discovering relationships
between products
4Jure Leskove, Pinterest & Stanford University
5. Product Graph
Ingest product catalogs:
10s of millions of products
100s of millions of descriptions, reviews
Infer product networks with multiple
types of directed relationships:
Input:
Data about items (products)
Output:
Network with multiple types of relationships
5Jure Leskove, Pinterest & Stanford University
7. Product Graph: Description
7
: cleaner; quieter
: cheaper; high power
: well made, easy to install
: fits perfectly, great value
Jure Leskove, Pinterest & Stanford University
9. Product Graph:What it does?
1. Understands the notions of
substitute and complement goods
is substitutable for
complements
9Jure Leskove, Pinterest & Stanford University
10. Product Graph:What it does?
2. Generates explanations
of why certain products are
preferred
“Good quality, soft, light
weight, the colors are
beautiful and exactly like
the picture!”
People prefer this
because:
10Jure Leskove, Pinterest & Stanford University
11. Product Graph:What it does?
3. Discovers micro-categories
of products
Small clusters of tightly related products:
11Jure Leskove, Pinterest & Stanford University
12. Product Graph:What it does?
4. Recommends
baskets of related products
Query: Suggested outfit:
Query: Suggested outfit:
12Jure Leskove, Pinterest & Stanford University
13. Product Graph: Overview
Building networks from products
Modeling: Can we use product data
to model product relationships?
Understanding: Can we explain
why people prefer certain products
over others?
13Jure Leskove, Pinterest & Stanford University
14. Problem Setting
Binary prediction task:
Given a pair of products, x and y, predict
whether they are related
(substitute/complementary)
Goal: Build a probabilistic model
that encodes
14Jure Leskove, Pinterest & Stanford University
15. Problem Setting
How to learn
from data
Train by maximum likelihood:
15
XComplementary
Not
Complementary
Jure Leskove, Pinterest & Stanford University
16. Approach
Products are described by their properties:
Review text, Product description,
Brand, Price, …
[0.3, 0, 0, 0.3, 0, 0, 0.2, 0, 0, 0.1]
[0.1, 0, 0, 0, 0.2, 0, 0, 0.1, 0, 0.2]
Challenges:
How do we discover right features?
How do we explain relationships?
How do we identify micro-categories?
16
Shoes Female
Jure Leskove, Pinterest & Stanford University
17. Our Solution: SCEPTRE
Link Prediction Review “topics”
Discover topics that “explain” product relations
17
Learn to discover topics that
explain the product graph
Jure Leskove, Pinterest & Stanford University
18. Challenges: Relation Direction
why do people who view
X eventually buy Y?
Relationships we want to learn
are not symmetric
18
Relationships: Explained by product “properties”
“baby, pajamas, pants, colorful”
Directedness: Subjective/qualitative language
“true size, fits well, items are the same color as on the picture”
Jure Leskove, Pinterest & Stanford University
19. Challenges: Multiple Relations
19
We want to learn multiple
relationships simultaneously
Solution: Learn multiple regressors (one for each
graph), that operate on a single set of topics
Jure Leskove, Pinterest & Stanford University
20. Challenges: Micro-Categories
20
Model discovers thousands of topics
but no micro-categories
Solution: Product hierarchy
Laptop charger specific topics
are only active for chargers.
These are micro-categories.
Topics at the top are common to all
electronics products, and will contain
generic electronics language
Associate each node in the category
tree with a small number of topics:
Jure Leskove, Pinterest & Stanford University
21. Building the Graph
C++ implementation that runs
on a single (large-memory) machine
OpenMP to parallelize computations
Experimental results:
Active part of the Amazon catalog
10m products
150m reviews
250m relationships
21Jure Leskove, Pinterest & Stanford University
33. References
Inferring Networks of Substitutable and Complementary
Products by J. McAuley, R. Pandey, J. Leskovec. ACM SIGKDD
International Conference on Knowledge Discovery and Data
Mining (KDD), 2015.
Hidden Factors and Hidden Topics: Understanding Rating
Dimensions with Review Text by J. McAuley, J. Leskovec. ACM
Conference on Recommender Systems (RecSys), 2013.
Learning Attitudes and Attributes from Multi-Aspect
Reviews by J. McAuley, J. Leskovec, D. Jurafsky. IEEE
International Conference On Data Mining (ICDM), 2012.
34Jure Leskove, Pinterest & Stanford University