Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Recommender System Seminar

Pacmann AI
October 14, 2019
990

Recommender System Seminar

Pacmann AI

October 14, 2019
Tweet

Transcript

  1. 1. Background: Definition, Problem Construction and Possible Solutions. 2. SVD-like

    for Recommender System: Matrix Factorization to model matrix completion problem 3. Implicit and Explicit signal: How to model Explicit and Implicit Signal with SVD-like RecSys 4. Recommender System workflow: Estimate, Filtering, Ranking and Randomization 5. Recommender System Cases Content
  2. • Chris Anderson in “The Long Tail ◦ “We are

    leaving the age of information and entering the age of recommendation” • CNN Money, “The race to create a 'smart' Google”: ◦ “The Web, they say, is leaving the era of search and entering one of discovery. What's the difference? Search is what you do when you're looking for something. Discovery is when something wonderful that you didn't know existed, or didn't know how to ask for, finds you.” Source: compiled from Xavier Amatriain MLSS slides (2014) Background
  3. • Value of Recommendation ◦ Netflix: 2/3 of the movies

    watched are recommended ◦ Google News: recommendations generate 38% more click through ◦ Amazon: 35% sales from recommendations • Metrics which Affected by Recommendation ◦ Activity Metrics: Increase in user retention ◦ Financial Metrics: Increase in sales ◦ Product Activity: Increase on number of unique items bought Source: compiled from Xavier Amatriain MLSS slides (2014) Background
  4. • The “Recommender problem” ◦ Estimate a utility function that

    automatically predicts how a user will like an item. ◦ Based on: ▪ Past behavior ▪ Relations to other users ▪ Item similarity ▪ Context Source: compiled from Xavier Amatriain MLSS slides (2014) Background
  5. The “Recommender problem” • Let C be set of all

    users and let S be set of all possible recommendable items • Let u be a utility function measuring the usefulness of item s to user c, i.e., u : C X S→R, where R is a totally ordered set • For each user c є C, we want to choose items s є S that maximize u. Utility is usually represented by rating but can be any function Source: compiled from Xavier Amatriain MLSS slides (2014) Background
  6. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation a. For example, Content based filtering, recommend similar items. Product name embedding with Word2Vec SVD-like Recommendation
  7. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation SVD-like Recommendation
  8. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation a. Recommend items based on similarity: SVD-like Recommendation
  9. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation a. Pros: i. Can be used if you don’t have any transaction history in the beginning ii. Sometimes can beat popular items benchmark. b. Cons: i. Low diversity metrics 1. Buy (mie-ayam), next recommendation (mie-ayam) SVD-like Recommendation
  10. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation SVD-like Recommendation
  11. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Non-Personalized Recommendation SVD-like Recommendation arg max similarity(item1, item2)
  12. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 2. Personalized Recommendation SVD-like Recommendation arg max Probability(item, user)
  13. Before dwelling to SVD-like RecSys CF, let’s focus on other

    approach: 1. Personalized Recommendation a. Based on items, users, and items-users interaction SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)
  14. Matrix Factorization • Unfortunately we can’t user Matrix Factorization for

    sparse data SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)
  15. Two types of Feedback Explicit and Implicit Feedback Source: compiled

    from Li Yen Kuo slides, Implicit RecSys (2019)
  16. In rating systems, such as MovieLens and Allmusic, the value

    of an entry denotes the rating of the item given by the user. Ratings can explicitly reflect the preference of an individual. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)
  17. For instance, in a music podcast service, the value of

    an entry may denote the subscription. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)
  18. Or play count Explicit and Implicit Feedback Source: compiled from

    Li Yen Kuo slides, Implicit RecSys (2019)
  19. SVD-like model is so easy to fit any feedback type

    1. Explicit a. Rating (Netflix Competition) Change the loss function to RMSE, regression task. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)
  20. SVD-like model is so easy to fit any feedback type

    1. Explicit b. Like/Dislike (Netflix Now) Change the loss function to log-loss, sigmoid activation. Explicit and Implicit Feedback
  21. SVD-like model is so easy to fit any feedback type

    2. Implicit a. Read/Not Read (Quora) Change loss function to BPR Loss Explicit and Implicit Feedback
  22. SVD-like model is so easy to fit any feedback type

    2. Implicit b. Frequency of Buy (Retail) Change loss function to Regression loss with Alpha, Koren (2009) Explicit and Implicit Feedback
  23. Modeling 1. Build Recommender System based on Watch/Not a. Implicit

    model, maximize click through rate 2. Build Recommender System based on Rating a. Explicit model, maximize preference 3. Build Recommender System based on Churn a. Implicit model, maximize retention Recommender System Workflow Source: compiled from Xavier Amatriain MLSS slides (2014)
  24. Recommender System Workflow Ranking • Most recommendations are presented in

    a sorted list • Recommendation can be understood as a ranking problem • Popularity is the obvious baseline • Ratings prediction is a clear secondary data input that allows for personalization • Many other features can be added
  25. Randomization Randomization • We will focus on Diversity Metrics vs

    Accuracy Metrics. • More diverse recommendation will increase Netflix CTR • More accurate recommendation will increase Netflix CTR • Diversity and Accuracy are negatively correlated.
  26. Temporal Recommendation Task Item 1 Item 4 Item 2 Item

    1 Item 2 Item 3 Item 2 Item 1 Item 5 RNN
  27. Searching Task 60 Pseudo code BPR For K times: 1.

    A = Sample random label positive 2. B = Sample random label negative 3. Dist = Model predict distance(A,B) 4. Update, make the distance farther.
  28. Searching Task • Pointwise ◦ for each items, train a

    classifier / regressor on it to predict how relevant it is • Pairwise ◦ given a pair of documents, compare which one has the highest rank. • Listwise ◦ sort the entire list of documents ▪ Direct optimization of IR measures such as NDCG
  29. Cowok 1 Cowok 3 Cowok 2 Cewek A Cewek C

    Cewek B Cowok 4 Cewek D Matching Task
  30. Matching Task Cowok 1 Cewek C Cewek D Cewek A

    Cewek B Cewek A Cowok 4 Cowok 2 Cowok 1 Cowok 3
  31. Pacmann AI Classes Quality State of the Art of Machine

    Learning Research Practical Skills Theoretical Understanding > 50 institutions 400++ alumni 6 Classes in the past [email protected] https://pacmann.ai