590 likes | 722 Views
Transparent User Models for Personalization. Khalid El-Arini Carnegie Mellon University Joint work with: Ulrich Paquet, Ralf Herbrich , Jurgen Van Gael, Blaise Agüera y Arcas. Personalization is ubiquitous. Personalization is invaluable. YouTube : 72+ hours/minute of new video
E N D
Transparent User Models for Personalization Khalid El-Arini Carnegie Mellon University Joint work with: Ulrich Paquet, Ralf Herbrich, Jurgen Van Gael, BlaiseAgüera y Arcas
Personalization is invaluable. • YouTube: 72+hours/minute of new video • Facebook: 950 million+ users • Twitter: 400+ million tweets/day • Shopping: [1994]: 500K unique consumer goods sold in U.S. [2010]: Amazon alone offered 24 million. Keyword search is not enough.
“Basil…is not a neo-Nazi. Lukas…isnot a shadowy stalker. David…is not Korean. intent on giving them such labels.” - J. Zaslow, November 26, 2002
What recourse do we have? • “there's just one way to change its mind: outfox it.” - J. Zaslow, November 26, 2002 Can we do better?
We propose an alternative. Why am I getting this? Vegan? Really? Why? You behave like a vegan hipster You: • tweeted with #meatlessmonday • follow @WholeFoods • …
We propose an alternative. Why am I getting this? Goal: Achieve transparency via interpretable user features, learned from user activity You behave like a Brooklyn hipster
Badges Goal: Achieve transparency via interpretable user features, learned from user activity You behave like a Brooklyn hipster
Approach Model Experiments Summary
Define a vocabulary of badges … vegan Apple fanboy runner photographer Rich, interpretable and explainable
Define a vocabulary of badges • Identify exemplars How do I find vegans?
observed label Take advantage of howusers describe themselves
Most vegans don’t label themselves as “vegan” on Twitter… we want to infer the attributes of these users
Define a vocabulary of badges • Identify exemplars • Model characteristic behavior • Hashtags#meatlessmonday • RetweetsRT@WholeFoods
Approach Model Experiments Summary
Model sketch • We have no negative training examples. Use a generative model. • Actions can be explained by multiple badges, even for the same user. Noisy-or to combine badges. • How do we deal with user corrections? Observing a latent variable.
B badges i=1…B
i=1…B N users u=1…N
j=1…F i=1…B F actions j=1…F u=1…N
bi(u) Does user u have badgei? j=1…F i=1…B j=1…F u=1…N
bi(u) λi(u) j=1…F i=1…B Does user u have labelfor badge i in his profile? j=1…F u=1…N
bi(u) λi(u) j=1…F i=1…B aj(u) j=1…F u=1…N Has user u performed action j?
Does badge i explain action j? sij bi(u) λi(u) j=1…F i=1…B aj(u) j=1…F u=1…N
sij What’s the probability that a user with badge i performs action j? bi(u) wi(u) φij αφ j=1…F i=1…B βφ aj(u) j=1…F u=1…N
sij What is the background probability for each action? bi(u) wi(u) φij αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
sij noisy or: Can at least one of mybadges (or the background) explain it? bi(u) wi(u) φij αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
sij bi(u) λi(u) φij αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
Beta priors to control sparsity sij bi(u) λi(u) φij αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
αT βT αF βF Beta prior to encode low recall (e.g., 10%) γiT γiF sij bi(u) λi(u) φij Beta prior to encode high precision (e.g., 99.9%) αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
αω βω αT βT αη βη αF βF ηi ωi γiT γiF sij bi(u) λi(u) φij αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
Inference • Collapsed Gibbs sampler (with MH steps) sij bi(u) φij φbg
αω βω αT βT αη βη αF βF ηi ωi γiT γiF sij bi(u) λi(u) φij You behave like a vegan hipster. αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
αω βω αT βT αη βη αF βF ηi ωi γiT γiF sij bi(u) λi(u) φij You behave like a vegan hipster. αφ j=1…F i=1…B βφ φbg aj(u) j=1…F u=1…N
Approach Model Experiments Summary
Data description • Start with 7 million Twitter users • Manually define 31 sample badges by specifying labels
Data description • Start with 7 million Twitter users • Manually define 31 sample badges by specifying labels • Gather 2million tweets from August 2011 • Recall: actions are hashtagsand retweets Remove infrequent actions and inactive users, leaving us with: 75,880 users 32,030 actions
Badge statistics artist photographer country music fan book worm
Do all badges look this good? No, but most do.
Over-generalized wine lover
Overwhelmed Ruby on Rails
Inferred Apple fanboy badge Self-described Apple fanboys
Comparative Analysis • Compare to labeled LDA [Ramage+ 2009] • LDA extension where each document is labeled with multiple tags • One-to-one mapping between topics and tags • Document explained only by topics associated with its tags • Hold out random 10% of labels, treat as ground truth, and try to predict them
Rank of held-out labels Better predictive performance better