Netflix uses a variety of techniques to provide personalized recommendations to users. Some key aspects include: 1. Netflix recommendations are generated using both offline and online techniques. Offline techniques allow for more complex computations but results may become stale, while online techniques can respond quickly but have stricter time constraints. 2. Recommendations are generated using a variety of data sources and machine learning models, including SVD, RBMs, gradient boosted trees, and other techniques. Both the data and models are important for generating high quality recommendations. 3. Netflix tests recommendations using both offline and online A/B testing techniques. Offline testing is used to evaluate new models and ideas before launching online tests involving real users

1. Netflix Recommendations
Beyond the 5 Stars
ACM SF-‐Bay Area
October 22, 2012
Xavier Amatriain
Personalization Science and Engineering -
‐Netflix
@xamat

2. Outline
1. The Netflix Prize & the Recommendation
Problem
2. Anatomy of Netflix Personalization
3. Data & Models
4. And…
a) Consumer (Data) Science
b) Or Software Architectures

3. 3

4. What we were interested in:
 High quality recommendations
Proxy question:
 Accuracy in predicted rating
 Improve by 10% = $1million!
Results
• Top 2 algorithms still in
production
SVD
RBM

5. What about the final prize ensembles?
 Our offline studies showed they were too computationally
intensive to scale
 Expected improvement not worth the engineering effort
 Plus…. Focus had already shifted to other issues that
had more impact than rating prediction.
5

6. Change of focus
6
2006 2012

7. Anatomy of
Netflix
Personalization
Everything is a Recommendation

8. Everything is personalized
Note:
Recommendations
are per household,
not individual user
Rows
Ranking
8

9. Top 10
Personalization awareness
Diversity
All
9
Dad Dad&Mom Daughter Mom
All Daughter Son Mom
All?

10. Support for Recommendations
10
Social Support

11. Social Recommendations
11

12. Watch again & Continue Watching
12

13. Genres
13

14. Genre rows
 Personalized genre rows focus on user interest
 Also provide context and “evidence”
 Important for member satisfaction – moving personalized rows to top on
devices increased retention
 How are they generated?
 Implicit: based on user’s recent plays, ratings, & other interactions
 Explicit taste preferences
 Hybrid:combine the above
 Also take into account:
 Freshness - has this been shown before?
 Diversity– avoid repeating tags and genres, limit number of TV genres, etc.

15. Genres - personalization
15

16. Genres - personalization
16

17. 17
Genres- explanations

18. Genres- explanations
18

19. 19
Genres – user involvement

20. Genres – user involvement
20

21.  Displayed in
many different
contexts
 In response to
user actions/
context (search,
queue add…)
 More like… rows
Similars

22. Anatomy of a Personalization - Recap
 Everything is a recommendation: not only rating
prediction, but also ranking, row selection, similarity…
 We strive to make it easy for the user, but…
 We want the user to be aware and be involved in the
recommendation process
 Deal with implicit/explicit and hybrid feedback
 Add support/explanations for recommendations
 Consider issues such as diversity or freshness
22

23. Data
&
Models

24. Big Data @Netflix
 Plays:
 Almost 30M subscribers
 Ratings: 4M/day
 Searches: 3M/day
30M/day
 2B hours streamed in Q4
2011
 1B hours in June 2012
24

25. Smart Models
25
 Logistic/linear regression
 Elastic nets
 SVD and other MF models
 Restricted Boltzmann Machines
 Markov Chains
 Different clustering approaches
 LDA
 Association Rules
 Gradient Boosted Decision Trees
 …

26. SVD
X[n x m] = U[n x r] S  r x r] (V[m x r])T
 X: m x n matrix (e.g., m users, n videos)
 U: m x r matrix (m users, r concepts)
 S: r x r diagonal matrix (strength of each ‘concept’) (r: rank of the matrix)
 V: r x n matrix (n videos, r concepts)

27. Simon Funk’s SVD
 One of the most
interesting findings
during the Netflix
Prize came out of a
blog post
 Incremental, iterative,
and approximate way
to compute the SVD
using gradient
descent
27

28. SVD for Rating Prediction
 Where f
 qv, xv, yv   are three item factor vectors
 Users are not parametrized, but rather represented by:
 R(u): items rated by user u
 N(u): items for which the user has given implicit preference (e.g. rated vs. not rated)
f
pu  
f
qv  
r' T
 b  p q
uv uv u v
 User factor vectors and item-factors vector
 Baseline buv   bu bv (user & item deviation from average)
 Predict rating as
 SVD++ (Koren et. Al) asymmetric variation w. implicit feedback
r'
uj uj j
jR(u)

28

 b  qT
R(u)

2  (r  b )x  N(u) 2
uv uv v 
 1 1 

jN (u) 
y 
j 

29. Artificial Neural Networks – 4 generations
 1st - Perceptrons (~60s)
 Single layer of hand-coded features
 Linear activation function
 Fundamentally limited in what they can learn to do.
 2nd - Back-propagation (~80s)
 Back-propagate error signal to get derivatives for learning
 Non-linear activation function
 3rd - Belief Networks (~90s)
 Directed acyclic graph composed of (visible & hidden) stochastic variables
with weighted connections.
 Infer the states of the unobserved variables & learn interactions between
variables to make network more likely to generate observed data.
29

30. Restricted Boltzmann Machines
 Restrict the connectivity to make learning easier.
 Only one layer of hidden units.
 Although multiple layers are possible
 No connections between hidden units.
 Hidden units are independent given the visible
states..
 So we can quickly get an unbiased sample from
the posterior distribution over hidden “causes”
when given a data-vector
 RBMs can be stacked to form Deep Belief
Nets (DBN) – 4th generation of ANNs
i
hidden
j
visible

31. RBM for the Netflix Prize
31

32. Ranking Key algorithm, sorts titles in most
contexts

33. Ranking
 Ranking = Scoring + Sorting + Filtering
bags of movies for presentation to a user
 Goal: Find the best possible ordering of a
set of videos for a user within a specific
context in real-time
 Objective: maximize consumption
 Aspirations: Played & “enjoyed” titles have
best score
 Akin to CTR forecast for ads/search results
 Factors
 Accuracy
 Novelty
 Diversity
 Freshness
 Scalability
 …

34. Ranking
 Popularity is the obvious baseline
 Ratings prediction is a clear secondary data
input that allows for personalization
 We have added many other features (and tried
many more that have not proved useful)
 What about the weights?
 Based on A/B testing
 Machine-learned

35. Example: Two features, linear model
35
Popularity
Predicted
Rating
1
2
3
4
5
Linear Model:
frank(u,v) = w1 p(v) + w2 r(u,v) + b
Final
Ranking

36. Ranking

37. Ranking

38. Ranking

39. Ranking

40. Learning to rank
 Machine learning problem: goal is to construct ranking
model from training data
 Training data can have partial order or binary judgments
(relevant/not relevant).
 Resulting order of the items typically induced from a
numerical score
 Learning to rank is a key element for personalization
 You can treat the problem as a standard supervised
classification problem
40

41. Learning to Rank Approaches
1. Pointwise
 Ranking function minimizes loss function defined on individual
relevance judgment
 Ranking score based on regression or classification
 Ordinal regression, Logistic regression, SVM, GBDT, …
2. Pairwise
 Loss function is defined on pair-wise preferences
 Goal: minimize number of inversions in ranking
 Ranking problem is then transformed into the binary classification
problem
 RankSVM, RankBoost, RankNet, FRank…

42. Learning to rank - metrics
 Quality of ranking measured using metrics as
 Normalized Discounted Cumulative Gain
 Mean Reciprocal Rank (MRR)
 Fraction of Concordant Pairs (FCP)
 Others…
 But, it is hard to optimize machine-learned
models directly on these measures (they are
not differentiable)
 Recent research on models that directly
optimize ranking measures
42
NDCG 
DCG
IDCG
DCG  relevance1
relevance
log i
2 2
n
 i
H hH rank(hi )
1 1
MRR  
FCP  ij
CP(xi, xj )
n(n 1)
2

43. Learning to Rank Approaches
3. Listwise
a. Indirect Loss Function
 RankCosine: similarity between ranking list and ground truth as loss function
 ListNet: KL-divergence as loss function by defining a probability distribution
 Problem: optimization of listwise loss function may not optimize IR metrics
b. Directly optimizing IR measures (difficult since they are not differentiable)
 Directly optimize IR measures through Genetic Programming
 Directly optimize measures with Simulated Annealing
 Gradient descent on smoothed version of objective function (e.g. CLiMF
presented at Recsys 2012 or TFMAP at SIGIR 2012)
 SVM-MAP relaxes the MAP metric by adding it to the SVM constraints
 AdaRank uses boosting to optimize NDCG

44. 44
Similars
 Different similarities computed
from different sources: metadata,
ratings, viewing data…
 Similarities can be treated as
data/features
 Machine Learned models
improve our concept of “similarity”

45. Data & Models - Recap
 All sorts of feedback from the user can help generate better
recommendations
 Need to design systems that capture and take advantage of
all this data
 The right model is as important as the right data
 It is important to come up with new theoretical models, but
also need to think about application to a domain, and practical
issues
 Rating prediction models are only part of the solution to
recommendation (think about ranking, similarity…)
45

46. More data or better models?
46
Really?
Anand Rajaraman: Stanford & Senior VP at
Walmart Global eCommerce (former Kosmix)

47. Sometimes, it’s not
about more data
47
More data or better models?

48. [Banko and Brill, 2001]
48
Norvig: “Google does not
have better Algorithms,
only more Data”
Many features/
low-bias models
More data or better models?

49. 0.03
0.02
0.01
0
49
0.04
0.05
0.09
0.08
0.07
0.06
0 1000000 2000000 3000000 4000000 5000000 6000000
Model performance vs. sample size
(actual Netflix system)
More data or better models?
Sometimes, it’s not
about more data

50. 50
More data or better models?
Data without a sound approach = noise

51. Consumer
(Data) Science

52. Consumer Science
52
 Main goal is to effectively innovate for customers
 Innovation goals
 “If you want to increase your success rate, double
your failure rate.” – Thomas Watson, Sr., founder of
IBM
 The only real failure is the failure to innovate
 Fail cheaply
 Know why you failed/succeeded

53. Consumer (Data) Science
53
1. Start with a hypothesis:
 Algorithm/feature/design X will increase member engagement
with our service, and ultimately member retention
2. Design a test
 Develop a solution or prototype
 Think about dependent & independent variables, control,
significance…
3. Execute the test
4. Let data speak for itself

54. Offline/Online testing process
days Weeks to months
Rollout
Feature to
all users
Offline
testing
Online A/B
testing
[success] [success]
[fail]
54

55. Offline testing
55
 Optimize algorithms offline
 Measure model performance, using metrics such as:
 Mean Reciprocal Rank, Normalized Discounted Cumulative Gain, Fraction of
Concordant Pairs, Precision/Recall & F-measures, AUC, RMSE, Diversity…
 Offline performance used as an indication to make informed
decisions on follow-up A/B tests
 A critical (and unsolved) issue is how offline metrics can
correlate with A/B test results.
 Extremely important to define a coherent offline evaluation
framework (e.g. How to create training/testing datasets is not
trivial)

56. Executing A/B tests
56
 Many different metrics, but ultimately trust user
engagement (e.g. hours of play and customer retention)
 Think about significance and hypothesis testing
 Our tests usually have thousands of members and 2-20 cells
 A/B Tests allow you to try radical ideas or test many
approaches at the same time.
 We typically have hundreds of customer A/B tests running
 Decisions on the product always data-driven

57. What to measure
57
 OEC: Overall Evaluation Criteria
 In an AB test framework, the measure of success is key
 Short-term metrics do not always align with long term
goals
 E.g. CTR: generating more clicks might mean that our
recommendations are actually worse
 Use long term metrics such as LTV (Life time value)
whenever possible
 In Netflix, we use member retention

58. What to measure
 Short-term metrics can sometimes be informative, and
may allow for faster decision-taking
 At Netflix we use many such as hours streamed by users or
%hours from a given algorithm
 But, be aware of several caveats of using early decision
mechanisms
Initial effects appear to trend.
See “Trustworthy Online
Controlled Experiments: Five
Puzzling Outcomes
Explained” [Kohavi et. Al. KDD
12]
58

59. Consumer Data Science - Recap
59
 Consumer Data Science aims to innovate for the
customer by running experiments and letting data speak
 This is mainly done through online AB Testing
 However, we can speed up innovation by experimenting
offline
 But, both for online and offline experimentation, it is
important to chose the right metric and experimental
framework

60. Architectures
60

61. Technology
61
http://techblog.netflix.com

62. 62

63. Event & Data
Distribution
63

64. • UI devices should broadcast many
different kinds of user events
• Clicks
• Presentations
• Browsing events
• …
• Events vs. data
• Some events only need to be
propagated and trigger an action
(low latency, low information per
event)
• Others need to be processed and
“turned into” data (higher latency,
higher information quality).
• And… there are many in between
• Real-time event flow managed
through internal tool (Manhattan)
• Data flow mostly managed through
Hadoop.
Event & Data Distribution
64

65. Offline Jobs
65

66. • Two kinds of offline jobs
• Model training
• Batch offline computation of
recommendations/
intermediate results
• Offline queries either in Hive or
PIG
• Need a publishing mechanism
that solves several issues
• Notify readers when result of
query is ready
• Support different repositories
(s3, cassandra…)
• Handle errors, monitoring…
• We do this through Hermes
Offline Jobs
66

67. Computation
67

68. • Two ways of computing personalized
results
• Batch/offline
• Online
• Each approach has pros/cons
• Offline
+ Allows more complex computations
+ Can use more data
- Cannot react to quick changes
- May result in staleness
• Online
+ Can respond quickly to events
+ Can use most recent data
- May fail because of SLA
- Cannot deal with “complex”
computations
• It’s not an either/or decision
• Both approaches can be combined
68
Computation

69. Signals & Models
69

70. • Both offline and online algorithms are
based on three different inputs:
• Models: previously trained from
existing data
• (Offline) Data: previously
processed and stored information
• Signals: fresh data obtained from
live services
• User-related data
• Context data (session, date,
time…)
70
Signals & Models

71. Results
71

72. • Recommendations can be serviced
from:
• Previously computed lists
• Online algorithms
• A combination of both
• The decision on where to service the
recommendation from can respond to
many factors including context.
• Also, important to think about the
fallbacks (what if plan A fails)
• Previously computed lists/intermediate
results can be stored in a variety of
ways
• Cache
• Cassandra
• Relational DB
72
Results

73. Alerts and Monitoring
73
 A non-trivial concern in large-scale recommender
systems
 Monitoring: continuously observe quality of system
 Alert: fast notification if quality of system goes below a
certain pre-defined threshold
 Questions:
 What do we need to monitor?
 How do we know something is “bad enough” to alert

74. What to monitor
 Staleness
 Monitor time since last data update
Did something go
wrong here?
74

75. What to monitor
 Algorithmic quality
 Monitor different metrics by comparing what users do and what
your algorithm predicted they would do
75

76. What to monitor
 Algorithmic quality
 Monitor different metrics by comparing what users do and what
your algorithm predicted they would do
Did something go
wrong here?
76

77. What to monitor
 Algorithmic source for users
 Monitor how users interact with different algorithms
Algorithm X
New version
Did something go
wrong here?
77

78. When to alert
78
 Alerting thresholds are hard to tune
 Avoid unnecessary alerts (the “learn-to-ignore problem”)
 Avoid important issues being noticed before the alert happens
 Rules of thumb
 Alert on anything that will impact user experience significantly
 Alert on issues that are actionable
 If a noticeable event happens without an alert… add a new alert
for next time

79. Conclusions
79

80. The Personalization Problem
80
 The Netflix Prize simplified the recommendation problem
to predicting ratings
 But…
 User ratings are only one of the many data inputs we have
 Rating predictions are only part of our solution
 Other algorithms such as ranking or similarity are very important
 We can reformulate the recommendation problem
 Function to optimize: probability a user chooses something and
enjoys it enough to come back to the service

81. More data +
Better models +
81
More accurate metrics +
Better approaches & architectures
Lots of room for improvement!

82. Thanks!
We’re hiring!
Xavier Amatriain (@xamat)
xamatriain@netflix.com