Presentation is loading. Please wait.

Presentation is loading. Please wait.

Yucheng LowMultiple Domain User Personalization Deepak Agarwal Yahoo! Research Yucheng Low Carnegie Mellon University Alexander J. Smola Yahoo! Research.

Similar presentations


Presentation on theme: "Yucheng LowMultiple Domain User Personalization Deepak Agarwal Yahoo! Research Yucheng Low Carnegie Mellon University Alexander J. Smola Yahoo! Research."— Presentation transcript:

1 Yucheng LowMultiple Domain User Personalization Deepak Agarwal Yahoo! Research Yucheng Low Carnegie Mellon University Alexander J. Smola Yahoo! Research

2 Yucheng LowMultiple Domain User Personalization Information Flood

3 Yucheng LowMultiple Domain User Personalization Personalization 3 Golf ReaderTech. Reader Can we provide personalization to new users?

4 Yucheng LowMultiple Domain User Personalization Movies User 1 User 2 Impossible when you have only one domain. Best you can do is to have a good baseline. One Domain Cold-Start

5 Yucheng LowMultiple Domain User Personalization MoviesNewsMusic Possible when you have many domains. Multiple Domains Cold Start

6 Yucheng LowMultiple Domain User Personalization Personalization across all domain Combine tokens from all spaces ignoring the source domain User Reads Golf News Watches MTV Golf, Tiger, Music, Song Expand token space to include source domain Golf:1, Tiger:1, Music:2, Song:2 Your Favorite Personalization Algorithm

7 Yucheng LowMultiple Domain User Personalization Personalization across all domain Combine tokens from all spaces ignoring the source domain User Reads Golf News Watches MTV Golf, Tiger, Music, Song Expand token space to include source domain Golf:1, Tiger:1, Music:2, Song:2 Your Favorite Personalization Algorithm Domains with more observations will swamp out all other domains What is a good personalization algorithm that will work for all domains?

8 Yucheng LowMultiple Domain User Personalization Solution Meta-Profile User Meta Profile User Music Profile User Music Profile User News Profile User News Profile Isolates each domain: Prevents larger domains from swamping out smaller domains. Personalized News Personalized Music

9 Yucheng LowMultiple Domain User Personalization Solution Meta-Profile User Meta Profile User Music Profile User Music Profile User News Profile User News Profile User Movie Profile User Movie Profile Extensible: domains can be added/removed easily

10 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation Basketball NBA, hoop Train 3-point Topic 1 Golf, Tiger, Woods, Club, Green, Hole-in-one Topic 2 Machine, Learning, Neural, Network, Train Topic 3 Document Topic 1 Topic 2 Topic 3 Michael I. Jordan trains a Neural Network to play golf 2 Golf 3 Network

11 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation N Document 1.Each document has a mixture over topics 2.For each word in each document a)Draw a topic b)Draw a word from the topic A document is a bag of words. A topic is a mixture of words.

12 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation N Document 1.Each document has a mixture over topics 2.For each word in each document a)Draw a topic b)Draw a word from the topic A document is a bag of words. A topic is a mixture of words. Document

13 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation N Document 1.Each document has a mixture over topics 2.For each word in each document a)Draw a topic b)Draw a word from the topic A document is a bag of words. A topic is a mixture of words. Document Sample From:

14 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation N Document 1.Each document has a mixture over topics 2.For each word in each document a)Draw a topic b)Draw a word from the topic A document is a bag of words. A topic is a mixture of words. Topic 1: Basketball, Michael, Jordan Topic 2: Golf, Tiger, Woods, Club, Green Topic 3: Machine, Learning, Neural

15 Yucheng LowMultiple Domain User Personalization Latent Dirichlet Allocation N Document 1.Each document has a mixture over topics 2.For each word in each document a)Draw a topic b)Draw a word from the topic A document is a bag of words. A topic is a mixture of words. Topics which make up each document Words which make up each topic

16 Yucheng LowMultiple Domain User Personalization Single Domain Personalization N 1.Each user has a mixture over topics 1.For each word in each document a)Draw a topic b)Draw a word from the topic A users interaction with a domain is a bag of words. A topic is a mixture of words. User Words which make up each topic Topics each user is interested in

17 Yucheng LowMultiple Domain User Personalization Multiple Domain Personalization N User us interaction with domain d User A users interaction with a domain is a bag of words. A topic is a mixture of words. Each user has a meta-profile: Each domain has a latent matrix: Users prior interest in a domain is

18 Yucheng LowMultiple Domain User Personalization Solution Meta-Profile User Meta Profile User Music Profile User Music Profile User News Profile User News Profile User Movie Profile User Movie Profile

19 Yucheng LowMultiple Domain User Personalization Users Music News Movies Topic->word table

20 Yucheng LowMultiple Domain User Personalization Gibbs Sampling N User us interaction with domain p LDA

21 Yucheng LowMultiple Domain User Personalization Gibbs Sampling N User us interaction with domain p Hold Constant Sample using LDA Sampler 1: Sample Hold Constant

22 Yucheng LowMultiple Domain User Personalization Gibbs Sampling N User us interaction with domain p Hold Constant 1: Sample 2: Sample Sample Langevin Diffusion

23 Yucheng LowMultiple Domain User Personalization Gibbs Sampling N User us interaction with domain p Optimize Hold Constant 1: Sample 2: Sample 3: Optimize Hold Constant LBFGS

24 Yucheng LowMultiple Domain User Personalization Experiments

25 Yucheng LowMultiple Domain User Personalization Experiments @ Yahoo! 2 domain dataset. Frontpage and News clicks of 5.6 million users. Frontpage/News: Article text for each click. 3 domain dataset: Frontpage, News and MyYahoo clicks of 5.6 million users. MyYahoo: Only has article IDs for each click with no text. Not semantically meaningful. All user information was anonymized.

26 Yucheng LowMultiple Domain User Personalization Test Protocol Holdout proportion of users who see more than one domain. Hide one of those domain and try to predict the words. Prediction metric is cosine similarity Baseline is mean prediction.

27 Yucheng LowMultiple Domain User Personalization Implementation Distributed implementation in C++ using Memcached for communication. Alex Smola, Shravan Narayanamurthy An Architecture for Parallel Topic Models VLDB 2010 Distributed LBFGS line search: Implement standard MPI-like in Memcached. Broadcast Reduce Barrier Takes 2-3 days for 500 iterations on 30 machines

28 Yucheng LowMultiple Domain User Personalization 2 Property Sanity Check

29 Yucheng LowMultiple Domain User Personalization 2 Property

30 Yucheng LowMultiple Domain User Personalization Topic Examples FrontPage topics: ahmadinejad, official, bold, big, 11, version, president, attacks, statement, mahmoud, iran, lie, sept, makes, spark, bigscreen, epic, office, wins, expert, days, expert, made, quake, chile, shorter, nasa, axis, planets, entire, shifted, rotation, 88, changed, temblor, earths, greek, bug, transforms, News Topics: ap, toyota, acceleration, data, unintended, review, similar, access, nationwide, sudden, explain, devices, lawsuits, auto, boxes, press, interviews, crash, blamed, experts, htc, apple, android, iphone, patents, maker, google, related, phone, patent, lawsuit, suing, handset, nexus, taiwans, apples, violated, filed, 20, phones,

31 Yucheng LowMultiple Domain User Personalization 3 Property

32 Yucheng LowMultiple Domain User Personalization 3 Property

33 Yucheng LowMultiple Domain User Personalization sandra, oscar, oscars, red, carpet, bullock, golden, gown, bullocks, nominee, bestactress, sparkles, stunning, vienna, bachelor, jake, pavelka, giraldi, finale, show, stars, dancing, love, season, time, abc, bacteria, fight, super, struggling, developed, doctors, resistant, lethal, virtually, drugs, antibiotic, competitors, chad, film, movie, movies, films, director, story, avatar, james, time, hollywood, big, make, hes, star, Frontpage -> News Celebrity Entertainment Science Science Fiction

34 Yucheng LowMultiple Domain User Personalization iphone, apple, app, apps, ipod, google, store, apples, android, mac, mobile, touch, ipad, device, phone, college, year, earn, years, 000, bestpaid, average, 129, colleges, graduates, ten, alums, schools, actor, likes, health, care, bill, obama, president, rep, house, republican, senate, news, sen, democrats, fox, congress, reform drafts, player, nfl, scouts, team, riskiest, peril, bryant, dez, pick, talented, nba, james, news, News -> Frontpage home, bank, facing, ceo, gomez, eviction, rosalina, bought, cleaning, foreclosed, client, janitor, offices, surprising, video,, captured, inside, mountain, terrorist, observers, impresses, alqaidas, complexity, base, features, hideout, size, special, secret, struck,, PoliticsDevices College

35 Yucheng LowMultiple Domain User Personalization Extension User Meta Profile User Music Profile User Music Profile User News Profile User News Profile Latent Dirichlet Allocation Latent Dirichlet Allocation User Movie Profile User Movie Profile Latent Dirichlet Allocation

36 Yucheng LowMultiple Domain User Personalization Extension User Meta Profile User Music Profile User Music Profile User News Profile User News Profile Flexible: Allows different algorithm for each domain Linear Model Matrix Factorization User Movie Profile User Movie Profile fLDA

37 Yucheng LowMultiple Domain User Personalization It Is How You Use It User Meta Profile User Music Profile User Music Profile Personalized with Algorithm X Use the Meta Profile for Initialization.

38 Yucheng LowMultiple Domain User Personalization It Is How You Use It User Meta Profile User Music Profile User Music Profile Personalized with Algorithm X After initialization, Meta Profile has no influence on the profile.

39 Yucheng LowMultiple Domain User Personalization It Is How You Use It User Meta Profile User Music Profile User Music Profile Personalized with Algorithm X Periodically Update the Meta Profile and Domain Latent Matrix

40 Yucheng LowMultiple Domain User Personalization Conclusion An generic, extensible model for combining domain personalization schemes. Scalable inference procedure that extends to millions of users. Demonstrate strong predictive performance on a large real world data

41 Yucheng LowMultiple Domain User Personalization Questions?

42 Yucheng LowMultiple Domain User Personalization Joint Parameter Space Stuff User 1 User 2 User 3

43 Yucheng LowMultiple Domain User Personalization Joint Action Space Stuff Movies News Music User 1 User 2 User 3

44 Yucheng LowMultiple Domain User Personalization Poor Solution 1: Joint Parameter Space Combine tokens from all spaces ignoring the source domain. User Reads Golf News Watches MTV Golf + Music Features Entertainment + Sports News Entertainment + Sports TV Personalization ?

45 Yucheng LowMultiple Domain User Personalization Attempt 1: Joint Parameter Space Combine tokens from all spaces ignoring the source domain. User Reads Golf News Watches MTV Golf + Music Features Entertainment + Sports News Entertainment + Sports TV Personalization ? Larger domain will swamp out all observations from the smaller domains

46 Yucheng LowMultiple Domain User Personalization Attempt 2: Joint Action Space Expand token space to include source domain User Reads Golf News Watches MTV Golf:1 Tiger:1 Music:2 Song:2 Features Forces single personalization technique for all domains Total model complexitycomplexity of each domain >


Download ppt "Yucheng LowMultiple Domain User Personalization Deepak Agarwal Yahoo! Research Yucheng Low Carnegie Mellon University Alexander J. Smola Yahoo! Research."

Similar presentations


Ads by Google