Indigo Recommendations Project

1
Indigo Recommendations Project

• Arnold Binas, Laurent Charlin,
• Alex Levinshtein, Maksims Volkovs
• Artificial Intelligence Group
• University of Toronto

2
Presentation Outline

• Project Introduction
• Approaches
• Results
• Recommendation
• Visualization
• Whats the future of recommendations at Indigo?

3
Project Goals

• Book recommendation
• Friend recommendation

1 2 3
1 2 3
4
Input Data (I)

• Book ratings from chapters.indigo.cas members

5
Input Data (II)

• Members purchase history
• Book information (e.g. category)
• User community
• User information
• Virtual Bookshelves
• Reviews
• Recommendations
• Top-ten lists
• Existing friends

6
Presentation Outline

• Project Introduction
• Approaches
• Results
• Recommendation
• Visualization
• Whats the future of recommendations at Indigo

7
Challenge 1 People rate what they enjoy
8
Challenge 2 Many books, many users, but not
many ratings
9
Challenge 3

• Which information source should one use to make
  recommendations?
• Ratings? Shelves? Purchase History?
• Purchase/Shelves history
• Ratings are the most useful and expressive
  feedback from users
• Reviews might contain more useful information but
  having a computer understand them is difficult.

10
Book recommendations - Approach

• Goal Predict ratings for books that have not
  been rated
• Output the highest rated books

11
Rating prediction

• Method Collaborative Filtering a Machine
  learning techniqueUsers with similar
  tastes agree on new products

12
Collaborative Filtering - Under the scenes
UserDescriptor
BookDescriptor
I like 19th century Novels,
Political Biography
13
Modeling Ratings

• Given known ratings, learn
• User descriptors
• Book descriptors

14
Recommending books

• Predict the ratings
• Output n-highest rated books

1 2
3
15
Evaluation (Book Recommendation)

• Difference between known and predicted ratings
• High ratings for purchased / shelf books

16
Friend recommendations - approach

• Recommend people with similar interests
• Rate users based on interest similarity
• Output the highest rated users
• Similar ratings for books Similar
  interests
• Conclusion friend recommendations rely on
  rating prediction

17
Recall Modeling Ratings

• Given known ratings, learn
• User descriptors
• Book descriptors

Use these for user comparison
18
Recommending Friends
1 2
3

• Compare users
• Output similar users

Similar ?
19
Evaluation (Friend Recommendation)

• Performance on a manually ranked group of users
• High user ranking implies similar tastes in books
• Agreement in book rankings
• Agreement in purchased books
• Agreement in book shelves

20
Presentation Outline

• Project Introduction
• Approaches
• Results
• Recommendation
• Visualization
• Whats the future of recommendations at Indigo?

21
Presentation Outline

• Project Introduction
• Approaches
• Results
• Recommendation
• Visualization
• Whats the future of recommendations at Indigo?

22
Our recommendation ?

• Run A/B testing
• Figure out which members benefit from
  recommendations.
• Get more ratings.

23
Indigos data at a glance

• 300K ratings
• 26K users (with at least one rating)
• 87K products (with at least one rating)