ACM RecSys 2012: Recommender Systems, Today

Neal Lathia @Cambridge_Uni talk at @ds_dln #strataconf 02/10/12

Transcript

1. acm recsys 2012: recommender systems, today @neal_lathia
2. warning: daunting task lookout for twitter handles
3. why #recsys? information overload mailing lists; usenet news (1992) see: @jkonstan, @presnick
4. why #recsys? information overload filter failure movies; books; music (~1995)
5. why #recsys? information overload filter failure creating value advertising; engagement; connection (today)
6. @dtunkelang
7. (1) collaborative “based on the premise that people looking for information should be able to make use of what others have already found and evaluated” (maltz & ehrlick)
8. (2) query-less “in September 2010 Schmidt said that one day the combination of cloud computing and mobile phones would allow Google to pass on information to users without them even typing in search queries”
9. (3) discovery engines “we are leaving the age of information and entering the age of recommendation” (anderson)
12. input: ratings, clicks, views users → items process: SVD, kNN, RBM, etc. f(user, item) → prediction ~ rating output: prediction-ranked recommendations measure: |prediction – rating| (prediction – rating)2
13. traditional problems accuracy, scalability, distributed computation, similarity, cold-start, … (don't reinvent the wheel)
14. acm recsys 2012: 5 open problems
15. problem 1: predictions temporality, multiple co-occurring objectives: diversity, novelty, freshness, serendipity, explainability
17. problem 2: algorithms more algorithms vs. more data vs. more rating effort
18. what is your algorithm doing? f(user, item) → R f(user, item 1 , item 2 ) → R f(user, [item 1 ...item n ]) → R e.g., @alexk_z @abellogin
19. problem 3: users + ratings signals, context, groups, intents, interfaces
20. @xamat
21. problem 4: items lifestyle, behaviours, decisions, processes, software development
22. @presnick
23. problem 5: measurement ranking metrics vs. usability testing vs. A/B testing
24. Online Controlled Experiments: Introduction, Learnings, and Humbling Statistics http://www.exp-platform.com/Pages/2012RecSys.aspx
25. 3 key lessons
26. lesson 1: #recsys is an ensemble ...of disciplines statistics, machine learning, human-computer interaction, social network analysis, psychology
27. lesson 2: return to the domain user effort, generative models, cost of a freakommendation, value you seek to create
28. @plamere
29. lesson 3: join the #recsys community learn, build, research, deploy: @MyMediaLite, @LensKitRS @zenogantner, @elehack contribute, read: #recsyswiki, @alansaid
30. recommender systems, today @neal_lathia