1. SIGIR’13, July 28–August 1, 2013, Dublin, Ireland. Yu Yang 2016-03-06 Addressing Cold-Start in App Recommendation Latent User Models Constructed from Twitter Followers

2. Outline Introduction Problems Findings Approach Experiments Conclusions

3. Outline Introduction Problems Findings Approach Experiments Conclusions

4. Introduction users are able to access a substantial number of apps via App Stores CF can address this problem for apps that have sufficient ratings from past users CF does not have any user ratings to base recommendations on the cold-start problem

5. Introduction CF ineffective -- no prior ratings wait for sufficient ratings OR use CBF CBF similar to the user’s previously-consumed items lack of diversity

6. Introduction Method nascent information culled from Twitter use Twitter handles to access an app’s Twitter account extract the IDs of their Twitter-followers create pseudo-documents apply LDA to generate latent groups estimate the probability of a user liking the app

7. Outline Introduction Problems Findings Approach Experiments Conclusions

8. Problems two types of cold-start problems in CF in-matrix prediction out-of-matrix prediction

9. Problems Ⅰ in-matrix prediction items that have been rated by at least one user

10. Problems Ⅱ out-of-matrix prediction newly released items have never been rated e.g., items 4 and 5

11. Outline Introduction Problems Findings Approach Experiments Conclusions

12. Findings People use SNS to talk about many subjects— including apps the descriptions of some apps contain references to their Twitter accounts the “Angry Birds Star Wars” app has a line that says “follow @angrybirds on Twitter.”

13. Findings app mentions in SNS can precede formal user ratings in app stores In May 2012, Evernote’s Twitter account already had 120,000 followers and 1,300 tweets

14. Findings Can we merge information mined from the rich data in SNS to enhance the performance of app recommendation? can we address the cold-start weaknesses of the proprietary user models of recommender systems (e.g., the user profiles in an app store) by using nascent signals about apps from SNS (e.g., the user profiles in Twitter)?

15. Outline Introduction Problems Findings Approach Experiments Conclusions

16. Approach Apps and their Twitter-Followers Pseudo-Documents and Pseudo-Words Constructing Latent Groups Estimation of the Probability of How Likely the Target User Will Like the App

17. Approach Apps and their Twitter-Followers

18. Approach Pseudo-Documents and Pseudo-Words

19. Approach Constructing Latent Groups define LDA as a generative process that creates a series of tuples

20. Approach Constructing Latent Groups The LDA model is used to compute the probability that the presence of a Twitter-follower t indicates that it is “liked” (+) (or “disliked” (−)) by user u, given user u’s past interaction Tu and the learned parameters α and β.

21. Approach Constructing Latent Groups

22. Approach Estimation of the Probability of How Likely the Target User Will Like the App

23. Approach Apps and their Twitter-Followers Pseudo-Documents and Pseudo-Words Constructing Latent Groups Estimation of the Probability of How Likely the Target User Will Like the App

24. Approach Estimation of the Probability of How Likely the Target User Will Like the App

25. Approach Estimation of the Probability of How Likely the Target User Will Like the App

26. Approach Estimation of the Probability of How Likely the Target User Will Like the App

27. Approach Estimation of the Probability of How Likely the Target User Will Like the App

28. Outline Introduction Problems Findings Approach Experiments Conclusions

29. Experiments

36. Outline Introduction Problems Findings Approach Experiments Conclusions

37. by incorporating information from Twitter, our approach overcomes the difficulty of cold-start app recommendation and significantly outperforms other state-of-the-art recommendation techniques by up to 33%

38. Thanks ！ Q&A