machine learning at netflix scale

48

Upload: aish-fenton

Post on 06-May-2015

332 views

Category:

Engineering


0 download

DESCRIPTION

Netflix is the world’s leading Internet television network with over 48 million members in more than 40 countries enjoying more than one billion hours of TV shows and movies per month, including original series. Netflix uses machine learning to deliver a personalized experience to each one of our 48 million users. In this talk you will hear about the machine learning algorithms that power almost every part of the Netflix experience, including some of our recent work on distributed Neural Networks on AWS GPUs. You will also get an insight into the innovation approach that includes offline experimentation and online AB testing. Finally, you will learn about the system architectures that enable all of this at a Netflix scale.

TRANSCRIPT

Page 1: Machine Learning at Netflix Scale
Page 2: Machine Learning at Netflix Scale

Machine Learning At Netflix Scale

Aish Fenton Manager - Research Engineering @aishfenton

Page 3: Machine Learning at Netflix Scale

Everything is a recommendation

Page 4: Machine Learning at Netflix Scale

4

Page 5: Machine Learning at Netflix Scale

Top Picks for Aish

Page 6: Machine Learning at Netflix Scale

Movies based on books

Page 7: Machine Learning at Netflix Scale

Because you watched Bob’s Burgers

Page 8: Machine Learning at Netflix Scale
Page 9: Machine Learning at Netflix Scale

Rank based on your taste

Ran

k ba

sed

on y

our

tast

e

Page 10: Machine Learning at Netflix Scale

75% of plays come from homepage

Page 11: Machine Learning at Netflix Scale

Back Story…

Page 12: Machine Learning at Netflix Scale
Page 13: Machine Learning at Netflix Scale

Proxy question: ▪ Accuracy in predicted rating ▪ Improve by 10% = $1million!

What we were interested in: ▪ High quality recommendations

predicted

actual

Page 14: Machine Learning at Netflix Scale

SVD RBMs

Top two results still used in production!

Page 15: Machine Learning at Netflix Scale

>

Page 16: Machine Learning at Netflix Scale

2006 2013

Page 17: Machine Learning at Netflix Scale

• > 44M members

• > 40 countries

• > 5B hours in Q3 2013

• Log 100B events/day

• 31.62% of peak US downstream traffic

Page 18: Machine Learning at Netflix Scale

Data and Models

Page 19: Machine Learning at Netflix Scale

▪ > 40M subscribers ▪ Ratings: ~5M/day ▪ Searches: >3M/day ▪ Plays: > 50M/day ▪ Streamed hours: o 5B hours in Q3 2013

Geo Info

Time

Impressions

Device Info

Metadata

Social

Ratings

Demographics

Member Behavior

Plays

Page 20: Machine Learning at Netflix Scale

Aish House of Cards

Latent User Vector

Latent Item Vector

Page 21: Machine Learning at Netflix Scale

3.53

RU

M

u1 u2 u3

m1 !m2!m3

House of Cards

Aish Aish

House of Cards

Page 22: Machine Learning at Netflix Scale
Page 23: Machine Learning at Netflix Scale

Mean Rating My Bias

Movie Bias

Interaction

Page 24: Machine Learning at Netflix Scale

Mean Rating My Bias

Movie Bias

Interaction

3.55 = 2.50 + -1.5 + 1.2 + pq

My rating for House of Cards

Page 25: Machine Learning at Netflix Scale

R3.53

U

M

u1 u2 u3

m1 !m2!m3

House of Cards

Aish

2.35

1.34

Time

T

t1 t2 t3 Time

Page 26: Machine Learning at Netflix Scale

▪ Matrix/Tensor Factorization ▪ Regression models (Logistic, Linear, Elastic nets) ▪ Factorization Machines ▪ Restricted Boltzmann Machines ▪ Markov Chains & other graph models ▪ Clustering / Topic Models ▪ Neural Networks ▪ Association Rules ▪ GBDT/RF ▪ …

Page 27: Machine Learning at Netflix Scale

Popularity

+ Ratings

+ More Features & Optimized Models

0% 50%

100%

150%

200%

250%

300%

Improvement Over Baseline

Page 28: Machine Learning at Netflix Scale

Anatomy of a Machine Learning

Platform

Page 29: Machine Learning at Netflix Scale

Problem

Data

Experiment Offline

Produce Model

Test / Metrics

Page 30: Machine Learning at Netflix Scale

Near-line

Online

UI Clients

Event Distribution

Online Algs

Model Trainer

Pre-compute

AB Test Metrics

API Layer

Monitoring

Offline

Hadoop / Data Warehouse

Experimentation Platform

S3 / HDFS

Offline MetricsQuery Tools

Models

Models

Page 31: Machine Learning at Netflix Scale

Near-line

Online

UI Clients

Event Distribution

Online Algs

Model Trainer

Pre-compute

AB Test Metrics

API Layer

Monitoring

Offline

Hadoop / Data Warehouse

Experimentation Platform

S3 / HDFS

Offline MetricsQuery Tools

Models

Models

Page 32: Machine Learning at Netflix Scale

▪ App Logs ▪ User Actions

▪ Ratings ▪ Plays ▪ Queue Adds

▪ Algo Actions ▪ Impressions (Presentation Bias)

▪ Context ▪ Device Info ▪ User Demographics ▪ Social ▪ Time

▪ …

Many different types of data…

Page 33: Machine Learning at Netflix Scale
Page 34: Machine Learning at Netflix Scale
Page 35: Machine Learning at Netflix Scale

Near-line

Online

UI Clients

Event Distribution

Online Algs

Model Trainer

Pre-compute

AB Test Metrics

API Layer

Monitoring

Offline

Hadoop / Data Warehouse

Experimentation Platform

S3 / HDFS

Offline MetricsQuery Tools

Models

Models

Embedded

Embedded

Page 36: Machine Learning at Netflix Scale

Weights

Real-time popularity of movie

Page 37: Machine Learning at Netflix Scale

Example: Neural Network Training

Page 38: Machine Learning at Netflix Scale
Page 39: Machine Learning at Netflix Scale

θ

Input OutputHidden Layer

Page 40: Machine Learning at Netflix Scale

Input OutputHidden Layers

Page 41: Machine Learning at Netflix Scale

Neural Network Training

1,536 cores

G2 Instances $0.60 p/h

Page 42: Machine Learning at Netflix Scale

But… things can go astray

Page 43: Machine Learning at Netflix Scale
Page 44: Machine Learning at Netflix Scale
Page 45: Machine Learning at Netflix Scale

Near-line

Online

UI Clients

Event Distribution

Online Algs

Model Trainer

Pre-compute

AB Test Metrics

API Layer

Monitoring

Offline

Hadoop / Data Warehouse

Experimentation Platform

S3 / HDFS

Offline MetricsQuery Tools

Models

Models

Page 46: Machine Learning at Netflix Scale

RU

MPre-compute

u1 u2 u3Online

Page 47: Machine Learning at Netflix Scale

Near-line

Online

UI Clients

Event Distribution

Online Algs

Model Trainer

Pre-compute

AB Test Metrics

API Layer

Monitoring

Offline

Hadoop / Data Warehouse

Experimentation Platform

S3 / HDFS

Offline MetricsQuery Tools

Models

Models

Aish played HoC

Publish new model for Aish

Page 48: Machine Learning at Netflix Scale

Aish Fenton @aishfenton https://www.linkedin.com/profile/view?id=47917219