1. 1 ©MapR Technologies 2013 Which Algorithms Really Matter?

2. 2 ©MapR Technologies 2013 Me, Us  Ted Dunning, Chief Application Architect, MapR Committer PMC member, Mahout, Zookeeper, Drill Bought the beer at the first HUG  MapR Distributes more open source components for Hadoop Adds major technology for performance, HA, industry standard API’s  Info Hash tag - #mapr See also - @ApacheMahout @ApacheDrill @ted_dunning and @mapR

3. 3 ©MapR Technologies 2013 Which Learning Algorithms Really Matter? The set of algorithms that matter theoretically is different from the ones that matter commercially. Commercial importance often hinges on ease of deployment, robustness against perverse data and conceptual simplicity. Often, even accuracy can be sacrificed against these other goals. Commercial systems also often live in a highly interacting environment so off-line evaluations may have only limited applicability. I will describe several commercially important algorithms such as Thompson sampling (aka Bayesian Bandits), result dithering, on-line clustering and distribution sketches and will explain what makes these algorithms important in industrial settings.

4. 4 ©MapR Technologies 2013 Topic For Today  What is important? What is not?  Why?  What is the difference from academic research?  Some examples

5. 5 ©MapR Technologies 2013 What is Important?  Deployable  Robust  Transparent  Skillset and mindset matched?  Proportionate

6. 6 ©MapR Technologies 2013 What is Important?  Deployable – Clever prototypes don’t count if they can’t be standardized  Robust  Transparent  Skillset and mindset matched?  Proportionate

7. 7 ©MapR Technologies 2013 What is Important?  Deployable – Clever prototypes don’t count  Robust – Mishandling is common  Transparent – Will degradation be obvious?  Skillset and mindset matched?  Proportionate

8. 8 ©MapR Technologies 2013 What is Important?  Deployable – Clever prototypes don’t count  Robust – Mishandling is common  Transparent – Will degradation be obvious?  Skillset and mindset matched? – How long will your fancy data scientist enjoy doing standard ops tasks?  Proportionate – Where is the highest value per minute of effort?

9. 9 ©MapR Technologies 2013 Academic Goals vs Pragmatics  Academic goals – Reproducible – Isolate theoretically important aspects – Work on novel problems  Pragmatics – Highest net value – Available data is constantly changing – Diligence and consistency have larger impact than cleverness – Many systems feed themselves, exploration and exploitation are both important – Engineering constraints on budget and schedule

10. 10 ©MapR Technologies 2013 Example 1: Making Recommendations Better

11. 11 ©MapR Technologies 2013 Recommendation Advances  What are the most important algorithmic advances in recommendations over the last 10 years?  Cooccurrence analysis?  Matrix completion via factorization?  Latent factor log-linear models?  Temporal dynamics?

12. 12 ©MapR Technologies 2013 The Winner – None of the Above  What are the most important algorithmic advances in recommendations over the last 10 years? 1. Result dithering 2. Anti-flood

13. 13 ©MapR Technologies 2013 The Real Issues  Exploration  Diversity  Speed  Not the last fraction of a percent

14. 14 ©MapR Technologies 2013 Result Dithering  Dithering is used to re-order recommendation results – Re-ordering is done randomly  Dithering is guaranteed to make off-line performance worse  Dithering also has a near perfect record of making actual performance much better

15. 15 ©MapR Technologies 2013 Result Dithering  Dithering is used to re-order recommendation results – Re-ordering is done randomly  Dithering is guaranteed to make off-line performance worse  Dithering also has a near perfect record of making actual performance much better “Made more difference than any other change”

16. 16 ©MapR Technologies 2013 Simple Dithering Algorithm  Generate synthetic score from log rank plus Gaussian  Pick noise scale to provide desired level of mixing  Typically  Oh… use floor(t/T) as seed

17. 17 ©MapR Technologies 2013 Example … ε = 0.5

18. 18 ©MapR Technologies 2013 Example … ε = log 2 = 0.69

19. 19 ©MapR Technologies 2013 Exploring The Second Page

20. 20 ©MapR Technologies 2013 Lesson 1: Exploration is good

21. 21 ©MapR Technologies 2013 Example 2: Bayesian Bandits

22. 22 ©MapR Technologies 2013 Bayesian Bandits  Based on Thompson sampling  Very general sequential test  Near optimal regret  Trade-off exploration and exploitation  Possibly best known solution for exploration/exploitation  Incredibly simple

23. 23 ©MapR Technologies 2013 Thompson Sampling  Select each shell according to the probability that it is the best  Probability that it is the best can be computed using posterior  But I promised a simple answer

24. 24 ©MapR Technologies 2013 Thompson Sampling – Take 2  Sample θ  Pick i to maximize reward  Record result from using i

25. 25 ©MapR Technologies 2013 Fast Convergence

26. 26 ©MapR Technologies 2013 Thompson Sampling on Ads An Empirical Evaluation of Thompson Sampling - Chapelle and Li, 2011

27. 27 ©MapR Technologies 2013 Bayesian Bandits versus Result Dithering  Many useful systems are difficult to frame in fully Bayesian form  Thompson sampling cannot be applied without posterior sampling  Can still do useful exploration with dithering  But better to use Thompson sampling if possible

28. 28 ©MapR Technologies 2013 Lesson 2: Exploration is pretty easy to do and pays big benefits.

29. 29 ©MapR Technologies 2013 Example 3: On-line Clustering

30. 30 ©MapR Technologies 2013 The Problem  K-means clustering is useful for feature extraction or compression  At scale and at high dimension, the desirable number of clusters increases  Very large number of clusters may require more passes through the data  Super-linear scaling is generally infeasible

31. 31 ©MapR Technologies 2013 The Solution  Sketch-based algorithms produce a sketch of the data  Streaming k-means uses adaptive dp-means to produce this sketch in the form of many weighted centroids which approximate the original distribution  The size of the sketch grows very slowly with increasing data size  Many operations such as clustering are well behaved on sketches Fast and Accurate k-means For Large Datasets. Michael Shindler, Alex Wong, Adam Meyerson. Revisiting k-means: New Algorithms via Bayesian Nonparametrics. Brian Kulis, Michael Jordan.

32. 32 ©MapR Technologies 2013 An Example

33. 33 ©MapR Technologies 2013 An Example

34. 34 ©MapR Technologies 2013 The Cluster Proximity Features  Every point can be described by the nearest cluster – 4.3 bits per point in this case – Significant error that can be decreased (to a point) by increasing number of clusters  Or by the proximity to the 2 nearest clusters (2 x 4.3 bits + 1 sign bit + 2 proximities) – Error is negligible – Unwinds the data into a simple representation  Or we can increase the number of clusters (n fold increase adds log n bits per point, decreases error by sqrt(n)

35. 35 ©MapR Technologies 2013 Diagonalized Cluster Proximity

36. 36 ©MapR Technologies 2013 Lots of Clusters Are Fine

37. 37 ©MapR Technologies 2013 Typical k-means Failure Selecting two seeds here cannot be fixed with Lloyds Result is that these two clusters get glued together

38. 38 ©MapR Technologies 2013 Streaming k-means Ideas  By using a sketch with lots (k log N) of centroids, we avoid pathological cases  We still get a very good result if the sketch is created – in one pass – with approximate search  In fact, adaptive dp-means works just fine  In the end, the sketch can be used for clustering or …

39. 39 ©MapR Technologies 2013 Lesson 3: Sketches make big data small.

40. 40 ©MapR Technologies 2013 Example 4: Search Abuse

41. 41 ©MapR Technologies 2013 Recommendations Alice got an apple and a puppy Charles got a bicycle Alice Charles

42. 42 ©MapR Technologies 2013 Recommendations Alice got an apple and a puppy Charles got a bicycle Bob got an apple Alice Bob Charles

43. 43 ©MapR Technologies 2013 Recommendations What else would Bob like? ? Alice Bob Charles

44. 44 ©MapR Technologies 2013 Log Files Alice Bob Charles Alice Bob Charles Alice

45. 45 ©MapR Technologies 2013 History Matrix: Users by Items Alice Bob Charles ✔✔✔ ✔✔ ✔✔

46. 46 ©MapR Technologies 2013 Co-occurrence Matrix: Items by Items - - 12 11 1 1 2 1 How do you tell which co-occurrences are useful?. 0 0 0 0

47. 47 ©MapR Technologies 2013 Co-occurrence Binary Matrix 1 1 not 1

48. 48 ©MapR Technologies 2013 Indicator Matrix: Anomalous Co-Occurrence ✔ ✔ Result: The marked row will be added to the indicator field in the item document…

49. 49 ©MapR Technologies 2013 Indicator Matrix ✔ id: t4 title: puppy desc: The sweetest little puppy ever. keywords: puppy, dog, pet indicators: (t1) id: t4 title: puppy desc: The sweetest little puppy ever. keywords: puppy, dog, pet indicators: (t1) That one row from indicator matrix becomes the indicator field in the Solr document used to deploy the recommendation engine. Note: data for the indicator field is added directly to meta-data for a document in Solr index. You don’t need to create a separate index for the indicators.

50. 50 ©MapR Technologies 2013 Internals of the Recommender Engine 50

51. 51 ©MapR Technologies 2013 Internals of the Recommender Engine 51

52. 52 ©MapR Technologies 2013 Looking Inside LucidWorks What to recommend if new user listened to 2122: Fats Domino & 303: Beatles? Recommendation is “1710 : Chuck Berry” 52 Real-time recommendation query and results: Evaluation

53. 53 ©MapR Technologies 2013 Real-life example

54. 54 ©MapR Technologies 2013 Lesson 4: Recursive search abuse pays Search can implement recs Which can implement search

55. 55 ©MapR Technologies 2013 Summary

56. 56 ©MapR Technologies 2013

57. 57 ©MapR Technologies 2013 Me, Us  Ted Dunning, Chief Application Architect, MapR Committer PMC member, Mahout, Zookeeper, Drill Bought the beer at the first HUG  MapR Distributes more open source components for Hadoop Adds major technology for performance, HA, industry standard API’s  Info Hash tag - #mapr See also - @ApacheMahout @ApacheDrill @ted_dunning and @mapR