Slides from an introductory talk on machine learning, and why mathematicians should take interest in it.
This is a very basic introduction, for math undergraduates & other curious minds.
The Netflix prize: yet another million dollar problem
1. The Problem
Strategies
Some Funny New Science
The Netflix Prize:
yet another million dollar problem
David Bessis
Ecole Normale Sup´rieure, 27/01/2010
e
David Bessis The Netflix Prize: yet another million dollar problem
2. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
David Bessis The Netflix Prize: yet another million dollar problem
3. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
David Bessis The Netflix Prize: yet another million dollar problem
4. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
David Bessis The Netflix Prize: yet another million dollar problem
5. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Solutions must
David Bessis The Netflix Prize: yet another million dollar problem
6. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Solutions must
”be published in a refereed mathematics publication of
worldwide repute”
David Bessis The Netflix Prize: yet another million dollar problem
7. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Solutions must
”be published in a refereed mathematics publication of
worldwide repute”
”have general acceptance in the mathematics community two
years after”
David Bessis The Netflix Prize: yet another million dollar problem
8. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
David Bessis The Netflix Prize: yet another million dollar problem
9. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
David Bessis The Netflix Prize: yet another million dollar problem
10. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
David Bessis The Netflix Prize: yet another million dollar problem
11. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
David Bessis The Netflix Prize: yet another million dollar problem
12. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
David Bessis The Netflix Prize: yet another million dollar problem
13. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Applied Mathematics.
David Bessis The Netflix Prize: yet another million dollar problem
14. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Applied Mathematics Computer Science.
David Bessis The Netflix Prize: yet another million dollar problem
15. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Applied Mathematics Computer Science
Psychology.
David Bessis The Netflix Prize: yet another million dollar problem
16. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Applied Mathematics Computer Science
Psychology (do we really care?)
David Bessis The Netflix Prize: yet another million dollar problem
17. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Some Funny New Science.
David Bessis The Netflix Prize: yet another million dollar problem
18. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Some Funny New Science.
Clear rules.
David Bessis The Netflix Prize: yet another million dollar problem
19. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Some Funny New Science.
Reasonably clear rules.
David Bessis The Netflix Prize: yet another million dollar problem
20. The Problem
Strategies
Some Funny New Science
7 + 1 Million Dollar Problems
Millenium Prize Problems:
Funded in 2000 by the Clay Mathematical Institute.
Seven classical open problems in Mathematics.
Fuzzy rules.
The Poincar´ conjecture was solved by Perelman in 2003.
e
No award yet.
Netflix Prize:
Funded in 2006 by the DVD rental company Netflix.
A problem in Some Funny New Science.
Reasonably clear rules.
Prize awarded in September 2009.
David Bessis The Netflix Prize: yet another million dollar problem
21. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
David Bessis The Netflix Prize: yet another million dollar problem
22. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
David Bessis The Netflix Prize: yet another million dollar problem
23. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
David Bessis The Netflix Prize: yet another million dollar problem
24. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
Collaborative filtering: recommending products based on prior
evaluations by other users (just like Amazon does).
David Bessis The Netflix Prize: yet another million dollar problem
25. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
Collaborative filtering: recommending products based on prior
evaluations by other users (just like Amazon does).
The Netflix prize is a collaborative filtering competition:
David Bessis The Netflix Prize: yet another million dollar problem
26. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
Collaborative filtering: recommending products based on prior
evaluations by other users (just like Amazon does).
The Netflix prize is a collaborative filtering competition:
Based on a huge dataset of actual ratings by Netflix users.
David Bessis The Netflix Prize: yet another million dollar problem
27. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
Collaborative filtering: recommending products based on prior
evaluations by other users (just like Amazon does).
The Netflix prize is a collaborative filtering competition:
Based on a huge dataset of actual ratings by Netflix users.
Open to almost everyone.
David Bessis The Netflix Prize: yet another million dollar problem
28. The Problem
Rules
Strategies
Competition
Some Funny New Science
Context
Netflix has an “all-you-can-eat” pricing model.
They need their users to watch a lot of movies.
Beyond a few obvious choices, people don’t know what they
want to watch.
Collaborative filtering: recommending products based on prior
evaluations by other users (just like Amazon does).
The Netflix prize is a collaborative filtering competition:
Based on a huge dataset of actual ratings by Netflix users.
Open to almost everyone.
Endowed with a $1.000.000 prize.
David Bessis The Netflix Prize: yet another million dollar problem
29. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
David Bessis The Netflix Prize: yet another million dollar problem
30. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
The movie space M consists of 17 770 movies
(identified by integers 1, . . . , 17 770, and the associated list of titles
and release years is provided – this data is meaningful and minable).
David Bessis The Netflix Prize: yet another million dollar problem
31. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
The movie space M consists of 17 770 movies
(identified by integers 1, . . . , 17 770, and the associated list of titles
and release years is provided – this data is meaningful and minable).
The date space D spans the period Oct. 1998 – Dec. 2005
(extremely meaningful data; no time of day is provided).
David Bessis The Netflix Prize: yet another million dollar problem
32. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
The movie space M consists of 17 770 movies
(identified by integers 1, . . . , 17 770, and the associated list of titles
and release years is provided – this data is meaningful and minable).
The date space D spans the period Oct. 1998 – Dec. 2005
(extremely meaningful data; no time of day is provided).
The rating space R is {1, 2, 3, 4, 5} (”stars”).
David Bessis The Netflix Prize: yet another million dollar problem
33. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
The movie space M consists of 17 770 movies
(identified by integers 1, . . . , 17 770, and the associated list of titles
and release years is provided – this data is meaningful and minable).
The date space D spans the period Oct. 1998 – Dec. 2005
(extremely meaningful data; no time of day is provided).
The rating space R is {1, 2, 3, 4, 5} (”stars”).
The training dataset T contains 100 480 507 quadruples
(u, m, d, r ) ∈ U × M × D × R.
David Bessis The Netflix Prize: yet another million dollar problem
34. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Dataset
The user space U consists of 480 189 users
(identified by a meaningless non-sequential integral id).
The movie space M consists of 17 770 movies
(identified by integers 1, . . . , 17 770, and the associated list of titles
and release years is provided – this data is meaningful and minable).
The date space D spans the period Oct. 1998 – Dec. 2005
(extremely meaningful data; no time of day is provided).
The rating space R is {1, 2, 3, 4, 5} (”stars”).
The training dataset T contains 100 480 507 quadruples
(u, m, d, r ) ∈ U × M × D × R.
The qualifying dataset Q contains 2 817 131 triples
(u, m, d) ∈ U × M × D.
David Bessis The Netflix Prize: yet another million dollar problem
35. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone
David Bessis The Netflix Prize: yet another million dollar problem
36. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives
David Bessis The Netflix Prize: yet another million dollar problem
37. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives and
residents of Cuba, Iran, Syria, North Korea, Myanmar and Sudan.
David Bessis The Netflix Prize: yet another million dollar problem
38. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives and
residents of Cuba, Iran, Syria, North Korea, Myanmar and Sudan.
Participants can join efforts in teams.
David Bessis The Netflix Prize: yet another million dollar problem
39. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives and
residents of Cuba, Iran, Syria, North Korea, Myanmar and Sudan.
Participants can join efforts in teams.
They can upload their predictions up to once a day.
David Bessis The Netflix Prize: yet another million dollar problem
40. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives and
residents of Cuba, Iran, Syria, North Korea, Myanmar and Sudan.
Participants can join efforts in teams.
They can upload their predictions up to once a day.
Predictions are maps from the qualifying set Q to the interval
[1, 5].
David Bessis The Netflix Prize: yet another million dollar problem
41. The Problem
Rules
Strategies
Competition
Some Funny New Science
The Challenge
Open to everyone except Netflix employees and their relatives and
residents of Cuba, Iran, Syria, North Korea, Myanmar and Sudan.
Participants can join efforts in teams.
They can upload their predictions up to once a day.
Predictions are maps from the qualifying set Q to the interval
[1, 5].
The metric used to benchmark predictions is the RMSE (”root
of mean square error”)
1
RMSE = |predicted rating for q − actual rating for q|2
|Q|
q∈Q
David Bessis The Netflix Prize: yet another million dollar problem
42. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
David Bessis The Netflix Prize: yet another million dollar problem
43. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
Users tend to view and rate movies they like, so they typically
give 3, 4 or 5 stars rather than 1 or 2 (the above upper bound
is unrealistically pessimistic).
David Bessis The Netflix Prize: yet another million dollar problem
44. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
Users tend to view and rate movies they like, so they typically
give 3, 4 or 5 stars rather than 1 or 2 (the above upper bound
is unrealistically pessimistic).
A basic prediction consists of mapping a triple (u, m, d) to the
average rating obtained by the movie m.
David Bessis The Netflix Prize: yet another million dollar problem
45. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
Users tend to view and rate movies they like, so they typically
give 3, 4 or 5 stars rather than 1 or 2 (the above upper bound
is unrealistically pessimistic).
A basic prediction consists of mapping a triple (u, m, d) to the
average rating obtained by the movie m. It achieves 1.0540.
David Bessis The Netflix Prize: yet another million dollar problem
46. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
Users tend to view and rate movies they like, so they typically
give 3, 4 or 5 stars rather than 1 or 2 (the above upper bound
is unrealistically pessimistic).
A basic prediction consists of mapping a triple (u, m, d) to the
average rating obtained by the movie m. It achieves 1.0540.
At the beginning of the Challenge, Netflix’s in-house
prediction system Cinematch achieved 0.9514
(roughly a 10% improvement).
David Bessis The Netflix Prize: yet another million dollar problem
47. The Problem
Rules
Strategies
Competition
Some Funny New Science
Typical RMSEs
Theoretically, the RMSE cannot be greater than 2.
Users tend to view and rate movies they like, so they typically
give 3, 4 or 5 stars rather than 1 or 2 (the above upper bound
is unrealistically pessimistic).
A basic prediction consists of mapping a triple (u, m, d) to the
average rating obtained by the movie m. It achieves 1.0540.
At the beginning of the Challenge, Netflix’s in-house
prediction system Cinematch achieved 0.9514
(roughly a 10% improvement).
Netflix set the following target: obtain a further 10%
improvement over Cinematch.
David Bessis The Netflix Prize: yet another million dollar problem
48. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 1: a Cryptographic Trick
Netflix has secretly partitioned the qualifying set
Q = Q1 Q2
into two subsets of (approximately) equal sizes.
David Bessis The Netflix Prize: yet another million dollar problem
49. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 1: a Cryptographic Trick
Netflix has secretly partitioned the qualifying set
Q = Q1 Q2
into two subsets of (approximately) equal sizes.
The RMSE achieved on Q1 is revealed to participants
(there is a public leaderboard).
David Bessis The Netflix Prize: yet another million dollar problem
50. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 1: a Cryptographic Trick
Netflix has secretly partitioned the qualifying set
Q = Q1 Q2
into two subsets of (approximately) equal sizes.
The RMSE achieved on Q1 is revealed to participants
(there is a public leaderboard).
The RMSE achieved on Q2 is used to determine the winner.
David Bessis The Netflix Prize: yet another million dollar problem
51. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 1: a Cryptographic Trick
Netflix has secretly partitioned the qualifying set
Q = Q1 Q2
into two subsets of (approximately) equal sizes.
The RMSE achieved on Q1 is revealed to participants
(there is a public leaderboard).
The RMSE achieved on Q2 is used to determine the winner.
This prevented participants from “learning from the oracle”.
David Bessis The Netflix Prize: yet another million dollar problem
52. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 1: a Cryptographic Trick
Netflix has secretly partitioned the qualifying set
Q = Q1 Q2
into two subsets of (approximately) equal sizes.
The RMSE achieved on Q1 is revealed to participants
(there is a public leaderboard).
The RMSE achieved on Q2 is used to determine the winner.
This prevented participants from “learning from the oracle”.
The goal was to achieve 0.8572.
David Bessis The Netflix Prize: yet another million dollar problem
53. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
David Bessis The Netflix Prize: yet another million dollar problem
54. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders
David Bessis The Netflix Prize: yet another million dollar problem
55. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
David Bessis The Netflix Prize: yet another million dollar problem
56. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
The Challenge was to last for 30 more days after the goal was
achieved.
David Bessis The Netflix Prize: yet another million dollar problem
57. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
The Challenge was to last for 30 more days after the goal was
achieved.
The winner would be the team with the best RMSE after this
30 days period
David Bessis The Netflix Prize: yet another million dollar problem
58. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
The Challenge was to last for 30 more days after the goal was
achieved.
The winner would be the team with the best RMSE after this
30 days period (no backstabbing arXiv-style “I posted first” effect).
David Bessis The Netflix Prize: yet another million dollar problem
59. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
The Challenge was to last for 30 more days after the goal was
achieved.
The winner would be the team with the best RMSE after this
30 days period (no backstabbing arXiv-style “I posted first” effect).
Every detail was carefully anticipated (even the possibility of a
tie).
David Bessis The Netflix Prize: yet another million dollar problem
60. The Problem
Rules
Strategies
Competition
Some Funny New Science
Very Smart Rules 2: Crowd Psychology Tricks
The Challenged opened on October 2, 2006.
Annual $50.000 prizes were offered to current leaders provided
they made their current methodology public.
The Challenge was to last for 30 more days after the goal was
achieved.
The winner would be the team with the best RMSE after this
30 days period (no backstabbing arXiv-style “I posted first” effect).
Every detail was carefully anticipated (even the possibility of a
tie).
These smart rules, together with the $1.000.000 prize,
attracted thousands of participants.
David Bessis The Netflix Prize: yet another million dollar problem
61. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
David Bessis The Netflix Prize: yet another million dollar problem
62. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
October 2007: team KorBell leads with 0.8712 (8.43%
improvement).
David Bessis The Netflix Prize: yet another million dollar problem
63. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
October 2007: team KorBell leads with 0.8712 (8.43%
improvement).
October 2008: team “BellKor in BigChaos” (two teams
merging efforts) leads with 0.8616 (9.44% improvement).
David Bessis The Netflix Prize: yet another million dollar problem
64. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
October 2007: team KorBell leads with 0.8712 (8.43%
improvement).
October 2008: team “BellKor in BigChaos” (two teams
merging efforts) leads with 0.8616 (9.44% improvement).
June 26, 2009: the goal is achieved.
David Bessis The Netflix Prize: yet another million dollar problem
65. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
October 2007: team KorBell leads with 0.8712 (8.43%
improvement).
October 2008: team “BellKor in BigChaos” (two teams
merging efforts) leads with 0.8616 (9.44% improvement).
June 26, 2009: the goal is achieved.
July 26, 2009: Netflix stops gathering solutions.
David Bessis The Netflix Prize: yet another million dollar problem
66. The Problem
Rules
Strategies
Competition
Some Funny New Science
Timeline
October 2006: Cinematch RMSE = 0.9514.
October 2007: team KorBell leads with 0.8712 (8.43%
improvement).
October 2008: team “BellKor in BigChaos” (two teams
merging efforts) leads with 0.8616 (9.44% improvement).
June 26, 2009: the goal is achieved.
July 26, 2009: Netflix stops gathering solutions.
The winner is announced on September 18, 2009.
David Bessis The Netflix Prize: yet another million dollar problem
67. The Problem
Rules
Strategies
Competition
Some Funny New Science
The winning team
Three teams combined their results to win the competition:
BellKor
Bob Bell (AT&T)
Yehuda Koren (Yahoo)
Chris Volinsky (AT&T)
BigChaos
Michael Jahrer (Commendo research and consulting)
Andreas T¨scher (Commendo research and consulting)
o
Pragmatic Theory
Martin Chabbert (Pragmatic Theory)
Martin Piotte (Pragmatic Theory)
David Bessis The Netflix Prize: yet another million dollar problem
68. The Problem
Rules
Strategies
Competition
Some Funny New Science
The winning team
Three teams combined their results to win the competition:
BellKor
Bob Bell (AT&T)
Yehuda Koren (Yahoo)
Chris Volinsky (AT&T)
BigChaos
Michael Jahrer (Commendo research and consulting)
Andreas T¨scher (Commendo research and consulting)
o
Pragmatic Theory
Martin Chabbert (Pragmatic Theory)
Martin Piotte (Pragmatic Theory)
Their winnning submission achieved a RMSE of 0.8567 (10.06%
improvement over Cinematch.)
David Bessis The Netflix Prize: yet another million dollar problem
69. The Problem
Rules
Strategies
Competition
Some Funny New Science
The winning team
Three teams combined their results to win the competition:
BellKor
Bob Bell (AT&T)
Yehuda Koren (Yahoo)
Chris Volinsky (AT&T)
BigChaos
Michael Jahrer (Commendo research and consulting)
Andreas T¨scher (Commendo research and consulting)
o
Pragmatic Theory
Martin Chabbert (Pragmatic Theory)
Martin Piotte (Pragmatic Theory)
Their winnning submission achieved a RMSE of 0.8567 (10.06%
improvement over Cinematch.)
Another team, The Ensemble, achieved the same RMSE...
David Bessis The Netflix Prize: yet another million dollar problem
70. The Problem
Rules
Strategies
Competition
Some Funny New Science
The winning team
Three teams combined their results to win the competition:
BellKor
Bob Bell (AT&T)
Yehuda Koren (Yahoo)
Chris Volinsky (AT&T)
BigChaos
Michael Jahrer (Commendo research and consulting)
Andreas T¨scher (Commendo research and consulting)
o
Pragmatic Theory
Martin Chabbert (Pragmatic Theory)
Martin Piotte (Pragmatic Theory)
Their winnning submission achieved a RMSE of 0.8567 (10.06%
improvement over Cinematch.)
Another team, The Ensemble, achieved the same RMSE...
...and lost because their submission was posted 24 minutes later!
David Bessis The Netflix Prize: yet another million dollar problem
71. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
David Bessis The Netflix Prize: yet another million dollar problem
72. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
David Bessis The Netflix Prize: yet another million dollar problem
73. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
David Bessis The Netflix Prize: yet another million dollar problem
74. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
A triple (m, v , d) can be encoded on 7 bytes.
David Bessis The Netflix Prize: yet another million dollar problem
75. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
A triple (m, v , d) can be encoded on 7 bytes.
700 MB suffice to store the dataset.
David Bessis The Netflix Prize: yet another million dollar problem
76. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
A triple (m, v , d) can be encoded on 7 bytes.
700 MB suffice to store the dataset.
It is possible (necessary) to work in RAM.
David Bessis The Netflix Prize: yet another million dollar problem
77. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
A triple (m, v , d) can be encoded on 7 bytes.
700 MB suffice to store the dataset.
It is possible (necessary) to work in RAM.
Commodity hardware is sufficient.
David Bessis The Netflix Prize: yet another million dollar problem
78. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Computer implementation
Memory requirements:
Movies can be encoded on 2 bytes (17770 < 2562 ).
Viewers can be encoded on 3 bytes (480189 < 2563 ).
Dates can be encoded on 2 bytes.
A triple (m, v , d) can be encoded on 7 bytes.
700 MB suffice to store the dataset.
It is possible (necessary) to work in RAM.
Commodity hardware is sufficient.
(I have some Ruby code to interactively play with the dataset.)
David Bessis The Netflix Prize: yet another million dollar problem
79. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
David Bessis The Netflix Prize: yet another million dollar problem
80. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
David Bessis The Netflix Prize: yet another million dollar problem
81. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
David Bessis The Netflix Prize: yet another million dollar problem
82. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
David Bessis The Netflix Prize: yet another million dollar problem
83. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
David Bessis The Netflix Prize: yet another million dollar problem
84. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
David Bessis The Netflix Prize: yet another million dollar problem
85. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
Similarly, a user rated all the movies, and many just a few.
David Bessis The Netflix Prize: yet another million dollar problem
86. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
Similarly, a user rated all the movies, and many just a few.
Let F be the set of all final 9 ratings for all individual users.
David Bessis The Netflix Prize: yet another million dollar problem
87. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
Similarly, a user rated all the movies, and many just a few.
Let F be the set of all final 9 ratings for all individual users.
Then F = Q P, with P ⊂ T publicly tagged by Netflix.
David Bessis The Netflix Prize: yet another million dollar problem
88. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
Similarly, a user rated all the movies, and many just a few.
Let F be the set of all final 9 ratings for all individual users.
Then F = Q P, with P ⊂ T publicly tagged by Netflix.
Q is a random draw of 2/3 of F .
David Bessis The Netflix Prize: yet another million dollar problem
89. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Remarks
About 200 ratings per users.
This is likely caused by Cinematch’s data gathering procedure:
users sometime rate tens of movies on a single day.
This causes an insanely huge bias within the dataset (movies
are perceived differently when rated individually or within a
rating spree), not fully exploited by the winners.
Netflix, do you read me?
Some movies were rated by hundreds of thousands viewers,
some by just a few (long-tail distribution).
Similarly, a user rated all the movies, and many just a few.
Let F be the set of all final 9 ratings for all individual users.
Then F = Q P, with P ⊂ T publicly tagged by Netflix.
Q is a random draw of 2/3 of F .
Q resembles P but is very dissimilar from T .
David Bessis The Netflix Prize: yet another million dollar problem
90. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
David Bessis The Netflix Prize: yet another million dollar problem
91. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
Regressions.
David Bessis The Netflix Prize: yet another million dollar problem
92. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
Regressions.
Latent parameters methods (SVD).
David Bessis The Netflix Prize: yet another million dollar problem
93. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
Regressions.
Latent parameters methods (SVD).
Neural networks.
David Bessis The Netflix Prize: yet another million dollar problem
94. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
Regressions.
Latent parameters methods (SVD).
Neural networks.
SVM
David Bessis The Netflix Prize: yet another million dollar problem
95. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Algorithms
The machine learning toolbox consists of many methods:
Clustering methods.
Regressions.
Latent parameters methods (SVD).
Neural networks.
SVM
...
David Bessis The Netflix Prize: yet another million dollar problem
96. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
David Bessis The Netflix Prize: yet another million dollar problem
97. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
Think conceptually and discretely rather than globally and
continuously.
David Bessis The Netflix Prize: yet another million dollar problem
98. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
Think conceptually and discretely rather than globally and
continuously.
Put users and movies into categories (clustering introduces
unwanted discontinuities).
David Bessis The Netflix Prize: yet another million dollar problem
99. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
Think conceptually and discretely rather than globally and
continuously.
Put users and movies into categories (clustering introduces
unwanted discontinuities).
Learn from the probe.
David Bessis The Netflix Prize: yet another million dollar problem
100. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
Think conceptually and discretely rather than globally and
continuously.
Put users and movies into categories (clustering introduces
unwanted discontinuities).
Learn from the probe.
Dealing with 100 000 000 data isn’t a logic puzzle.
David Bessis The Netflix Prize: yet another million dollar problem
101. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Beginner’s mistakes
Underestimate the volume effect.
Think conceptually and discretely rather than globally and
continuously.
Put users and movies into categories (clustering introduces
unwanted discontinuities).
Learn from the probe.
Dealing with 100 000 000 data isn’t a logic puzzle.
It resembles Thermodynamics.
David Bessis The Netflix Prize: yet another million dollar problem
102. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
David Bessis The Netflix Prize: yet another million dollar problem
103. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
David Bessis The Netflix Prize: yet another million dollar problem
104. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
David Bessis The Netflix Prize: yet another million dollar problem
105. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
Write (rx,y ) = (Cy )y ∈Y where the Cy are the column vectors.
David Bessis The Netflix Prize: yet another million dollar problem
106. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
Write (rx,y ) = (Cy )y ∈Y where the Cy are the column vectors.
Performing the linear regression consists of approximating Cy0 by
ˆ
its orthogonal projection Cy0 on the hyperplane generated by the
(Cy )y ∈Y .
David Bessis The Netflix Prize: yet another million dollar problem
107. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
Write (rx,y ) = (Cy )y ∈Y where the Cy are the column vectors.
Performing the linear regression consists of approximating Cy0 by
ˆ
its orthogonal projection Cy0 on the hyperplane generated by the
(Cy )y ∈Y .
Clearly, there exists a unique solution.
David Bessis The Netflix Prize: yet another million dollar problem
108. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
Write (rx,y ) = (Cy )y ∈Y where the Cy are the column vectors.
Performing the linear regression consists of approximating Cy0 by
ˆ
its orthogonal projection Cy0 on the hyperplane generated by the
(Cy )y ∈Y .
Clearly, there exists a unique solution.
It optimizes RMSE.
David Bessis The Netflix Prize: yet another million dollar problem
109. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Linear regression
Suppose all viewers in X have rated all movies in Y : the rating
matrix is
(rx,y )(x,y )∈X ×Y .
Suppose you want to model the ratings given to a particular movie
y0 based on the ratings given to the movies in Y = Y − {y0 }.
A linear regression is a natural way to do that.
Write (rx,y ) = (Cy )y ∈Y where the Cy are the column vectors.
Performing the linear regression consists of approximating Cy0 by
ˆ
its orthogonal projection Cy0 on the hyperplane generated by the
(Cy )y ∈Y .
Clearly, there exists a unique solution.
It optimizes RMSE.
Write the formula!
David Bessis The Netflix Prize: yet another million dollar problem
110. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
David Bessis The Netflix Prize: yet another million dollar problem
111. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
David Bessis The Netflix Prize: yet another million dollar problem
112. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
David Bessis The Netflix Prize: yet another million dollar problem
113. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
David Bessis The Netflix Prize: yet another million dollar problem
114. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
Normalize ratings:
David Bessis The Netflix Prize: yet another million dollar problem
115. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
Normalize ratings:
replace the rating rv ,m by the meaningful signal, i.e., the difference
r v ,m between rv ,m and the average rating for m.
David Bessis The Netflix Prize: yet another million dollar problem
116. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
Normalize ratings:
replace the rating rv ,m by the meaningful signal, i.e., the difference
r v ,m between rv ,m and the average rating for m.
Then it becomes natural to set r v ,m to 0 when v hasn’t rated m.
David Bessis The Netflix Prize: yet another million dollar problem
117. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
Normalize ratings:
replace the rating rv ,m by the meaningful signal, i.e., the difference
r v ,m between rv ,m and the average rating for m.
Then it becomes natural to set r v ,m to 0 when v hasn’t rated m.
Actually, whether or not v has rated m is a meaningful information!
David Bessis The Netflix Prize: yet another million dollar problem
118. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 1: missing data
Not all viewers have seen all movies.
Worse, there are virtually no complete rectangular blocks
within the dataset.
Regression by viewers or by movies?
It is better to do regression by movies.
Normalize ratings:
replace the rating rv ,m by the meaningful signal, i.e., the difference
r v ,m between rv ,m and the average rating for m.
Then it becomes natural to set r v ,m to 0 when v hasn’t rated m.
Actually, whether or not v has rated m is a meaningful information!
Add normalized bit columns to account for that.
David Bessis The Netflix Prize: yet another million dollar problem
119. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
David Bessis The Netflix Prize: yet another million dollar problem
120. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
Similarly, the curse of dimensionality asserts that:
With high-dimensionality datasets, one will always find stupid
predictors, making perfect predictions on the dataset, and
failing to generalize.
David Bessis The Netflix Prize: yet another million dollar problem
121. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
Similarly, the curse of dimensionality asserts that:
With high-dimensionality datasets, one will always find stupid
predictors, making perfect predictions on the dataset, and
failing to generalize.
By looking at my audience today, what should I be able to infer?
David Bessis The Netflix Prize: yet another million dollar problem
122. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
Similarly, the curse of dimensionality asserts that:
With high-dimensionality datasets, one will always find stupid
predictors, making perfect predictions on the dataset, and
failing to generalize.
By looking at my audience today, what should I be able to infer?
That having long hair is a reasonably good gender predictor?
David Bessis The Netflix Prize: yet another million dollar problem
123. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
Similarly, the curse of dimensionality asserts that:
With high-dimensionality datasets, one will always find stupid
predictors, making perfect predictions on the dataset, and
failing to generalize.
By looking at my audience today, what should I be able to infer?
That having long hair is a reasonably good gender predictor?
That wearing a grey sweater is a reasonably good gender
predictor?
David Bessis The Netflix Prize: yet another million dollar problem
124. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Real life problems 2: the curse of dimensionality
We all know that Lagrange interpolators are not to be used on
noisy data. Rather, one should look at best-fitting polynomials of a
given low degree.
Similarly, the curse of dimensionality asserts that:
With high-dimensionality datasets, one will always find stupid
predictors, making perfect predictions on the dataset, and
failing to generalize.
By looking at my audience today, what should I be able to infer?
That having long hair is a reasonably good gender predictor?
That wearing a grey sweater is a reasonably good gender
predictor?
Dilemma: overlearning vs underlearning.
David Bessis The Netflix Prize: yet another million dollar problem
125. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Ridge regression (aka Tikhonov regularization)
Linear regression: given vectors x, y1 , . . . , yn ∈ Rm , find λ1 , . . . , λn
that minimize
||x − λi yi ||2 .
David Bessis The Netflix Prize: yet another million dollar problem
126. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Ridge regression (aka Tikhonov regularization)
Linear regression: given vectors x, y1 , . . . , yn ∈ Rm , find λ1 , . . . , λn
that minimize
||x − λi yi ||2 .
When n is large (with respect to m), the linear system is
overdetermined. Overfitting occurs.
David Bessis The Netflix Prize: yet another million dollar problem
127. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Ridge regression (aka Tikhonov regularization)
Linear regression: given vectors x, y1 , . . . , yn ∈ Rm , find λ1 , . . . , λn
that minimize
||x − λi yi ||2 .
When n is large (with respect to m), the linear system is
overdetermined. Overfitting occurs.
A telltale sign of overfitting is the presence of λi ’s with huge norms
compensating each other.
David Bessis The Netflix Prize: yet another million dollar problem
128. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Ridge regression (aka Tikhonov regularization)
Linear regression: given vectors x, y1 , . . . , yn ∈ Rm , find λ1 , . . . , λn
that minimize
||x − λi yi ||2 .
When n is large (with respect to m), the linear system is
overdetermined. Overfitting occurs.
A telltale sign of overfitting is the presence of λi ’s with huge norms
compensating each other.
Ridge regression (Tikhonov regularization): find λ1 , . . . , λn that
minimize
||x − λi yi ||2 + ε |λi |2
where ε is a well-adjusted (small) penalty term.
David Bessis The Netflix Prize: yet another million dollar problem
129. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Assigning attributes to movies
Assume that movies differ by their amount of certain qualities:
David Bessis The Netflix Prize: yet another million dollar problem
130. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Assigning attributes to movies
Assume that movies differ by their amount of certain qualities:
Violence.
David Bessis The Netflix Prize: yet another million dollar problem
131. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Assigning attributes to movies
Assume that movies differ by their amount of certain qualities:
Violence.
Sex.
David Bessis The Netflix Prize: yet another million dollar problem
132. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Assigning attributes to movies
Assume that movies differ by their amount of certain qualities:
Violence.
Sex.
Anything else?
David Bessis The Netflix Prize: yet another million dollar problem
133. Practical issues
The Problem
Regressions
Strategies
Latent factors
Some Funny New Science
Tuning and Blending
Assigning attributes to movies
Assume that movies differ by their amount of certain qualities:
Violence.
Sex.
Maybe not.
David Bessis The Netflix Prize: yet another million dollar problem