Content based Collaborative filtering Matrix factorization Recommender evaluation recall vs precision

2. Agenda
● What is a recommender system? And why is it a MUST?
● History
● Personalized Vs non personalized recommenders
● Content based filtering
● Collaborative filtering
● Recommender challenges
● How to evaluate a recommender
● Salesman for eCommerce
● Seeloz for retailer

3. What is a Recommender System?

4. Youtube

5. Amazon

6. Facebook

7. Facebook

8. Why Recommender System is a MUST?
● In 2012, Amazon reported a 29% sales increase to
$12.83 billion during its second fiscal quarter, up from
$9.9 billion during the same time last year mostly as a
result for its recommender system
● In 2012, 75% of Netflix content watched on the service
comes from its recommendation engine

9. Why recommender system is a MUST?
● Sell more diverse items
● Increase user purchase conversion rate
● Increase CTR click through rate
● Increase user loyalty/satisfaction
● Increase revenue

10. History
Netflix Prize
● 2006 - 2009
● $1,000,000
● Dataset of 100M movie ratings
● 10% more accurate than existing recommender

11. Recommend based on what?
● Product info
● User info
● Purchase history
● Ratings
● Reviews
● Likes
● Shares
● Clicks
● More user behaviour….

12. Non Personalized Recommender
● Highest rated
● Most sold
● Bought together

13. Non Personalized Recommender
● Highest rated problems:
⚪ Low confidence with few ratings
◾ Ex: Only one rating with 5
⚪ Old ratings

14. Personalized Recommender
● Content based filtering
● Collaborative filtering

15. Content Based Filtering
Profile
1. Item
⚪ Category, price, brand, ..etc
2. User
⚪ Age, gender, address, categories of interests..etc

16. Content Based Filtering
Action
Comedy
Die Hard
My taste

17. Content Based Limitations
● Very simple
⚪ I like comedies with violence, and historical
documentaries, but not historical comedies or
violent documentaries
● Figuring out the right weights and factors

18. Collaborative Filtering
● Neighborhood Methods
● Latent Factor Model

19. Collaborative Filtering - Neighborhood

20. Collaborative Filtering - Neighborhood

21. Collaborative Filtering Challenges
● Find similar users
● How many neighbors should I select?
● Sparse matrix
● Heavy computations (million of users, thousands of
items)
● Cold start
● Not real time (solution: use item-item)

22. Collaborative Filtering - Neighborhood
● Item-Item

23. Latent Factor Models
Finds the hidden
characteristics in data
Most successful
implementations are based
on Matrix Factorization

24. Matrix Factorization

26. 4 x 2 2 x 6 4 x 6
Matrix Factorization

27. Matrix Factorization

28. Matrix Factorization

29. Matrix Factorization

30. Challenges
● Diversity (Most of users watched only one movie)
● Scalability for large datasets
● Not only what to recommend but also WHEN and HOW
● Generic model
● High computations

31. How to Evaluate a Recommender System
● Precision (1 vs 0.80)
● Recall (0.5 vs 1)
● F-measure F=2∗P∗R/(P+R)
⚪ (0.6 vs 0.88)
● Advanced methods:
⚪ Normalized discounted
cumulative gain
⚪ Mean square error
⚪ AUC
User => A, B, F, G
Recommender 1: A, B
Recommender 2: A, B, C, F, G

32. Use cases
● eCommerce
⚪ Salesman
● Retail
⚪ Seeloz

35. Salesman Features
● Personalized
⚪ Recommended for you
⚪ Product related recommendations
⚪ Cart recommendations
● Non personalized
⚪ Top sold
⚪ Recently viewed
⚪ Bought/viewed together

36. Salesman Challenges
● Need a generic recommender
● Cold start
● Hybrid solution
● Track different user actions with dynamic weights
● Real time
● Big data

37. Seeloz for Retail

38. Seeloz Promotion Planning
● Promotion objectives
⚪ Lift
⚪ Shift
⚪ Retention
⚪ Profit max
⚪ Acquisition

39. Seeloz Challenges
● Integrating with different POS systems
● Products normalization
● Data cleaning
● Data aggregation
● Different promotion objectives
● Promotions performance monitoring
● Big data

40. Thanks!