82
Video Recommendation at Scale January 2017 RecSysFR Meetup
Video Recommendation at Scale
January 2017RecSysFR Meetup
Lucas CharrierSoftware Engineer
Thomas BelhalfaouiData Scientist
About Pulpix
About Pulpix
AI startup based in New-York and Paris
100+ sites using our technology in the world
$850,000 seed round raised in the US
Accelerated in Silicon Valley by
About Pulpix
AI startup based in New-York and Paris
100+ sites using our technology in the world
$850,000 seed round raised in the US
Accelerated in Silicon Valley by
About Pulpix
100+ sites using our technology in the world
Accelerated in Silicon Valley by
$850,000 seed round raised in the US
AI startup based in New-York and Paris
About Pulpix
100+ sites using our technology in the world
$850,000 seed round raised in the US
AI startup based in New-York and Paris
Accelerated in Silicon Valley by
What do we do at Pulpix?
Video recommendation
How?
People binge-watch videoson social platforms
Why?
UX
AI
There is one simple reason: they invest in UX and AI.
Media Website vs. Social Platform
Media Website Media Website
So, what do we do at Pulpix?
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
Video to Video
Video to Video
Article to Video
in Video
Recommendation at Scale
Recommendation at ScaleKey figures
10 million videos
Less than 100 ms response time
10 million events a day
More than one billion training events
Recommendation at ScaleKey figures
10 million videos
Less than 100 ms response time
10 million events a day
More than one billion events
Less than 100 ms response time
More than one billion training events
10 million events a day
10 million videos
Recommendation at ScaleKey figures
Less than 100 ms response time
10 million events a day
10 million videos
More than one billion training events
Recommendation at ScaleKey figures
From R&D to productionIterate fast
From R&D to productionIterate fast
Idea
Prototype
Retrospective evaluation
From R&D to productionIterate fast
Idea
Prototype
Retrospective evaluation
A/B test
24h to
10 days
Content-Based EngineFirst approach
Content-based EngineFirst approach
Speech-to-text
Metadata
Keywords extractionWeighting
Content-based EngineFirst approach
Speech-to-text
Content ScoreMetadata
Keywords extractionWeighting
Content-based EngineFirst approach
Speech-to-text
Content ScoreMetadata
Keywords extractionWeighting
Recency boost
Collaborative Filtering
Collaborative FilteringMatrix factorization
Videos
Collaborative FilteringMatrix factorization
Videos
Collaborative FilteringMatrix factorization
Implicit rating
ExampleL’Equipe.fr
ExampleL’Equipe.fr
Football Basket Tennis54% 12% 26%
User preferences
Collaborative FilteringHow to put it into practice?
• User-based recommendations- Known users only- Not contextual
Collaborative FilteringHow to put it into practice?
• User-based recommendations- Known users only- Not contextual
• Video-based recommendations- For all users- Fully contextual
Hybrid Engine
Hybrid engineKey dimensions
ContextHybrid engineKey dimensions
PersonalizationContextHybrid engineKey dimensions
Personalization RecencyContextHybrid engineKey dimensions
Hybrid EngineLinear score combination
ContentEngine
CollaborativeEngine
Context
Recency
Global scoreScores Linear
Model
Hybrid EngineNonlinear embedding combination
ContentEngine
CollaborativeEngine
Context
Recency
Scores Candidate videos
Hybrid EngineNonlinear embedding combination
ContentEngine
CollaborativeEngine
Context
Recency
Global score
Scores Candidate videos
NonlinearModel
Features
Scalable Data Science stack
Flexibility queuing
Scalability sharding
Fault tolerance replication
High throughput replication
Recommendation at ScaleRequirements
Flexibility queuing
Scalability sharding
Fault tolerance replication
High throughput replication
Recommendation at ScaleRequirements
Flexibility queuing
Scalability sharding
Fault tolerance replication
High throughput replication
Recommendation at ScaleRequirements
Flexibility queuing
Scalability sharding
Fault tolerance replication
High throughput replication
Recommendation at ScaleRequirements
Recommendation at ScaleOur Data Science stack
Recommendation at ScaleOur Data Science stack
tracking
Recommendation at ScaleOur Data Science stack
tracking ML
dataprocessing
Recommendation at ScaleOur Data Science stack
tracking
reco
ML
dataprocessing
What’s next?
• Reinforcement Learning
What’s next?Our current R&D
• Reinforcement Learning
• Deep Learning:
○ Recommendation
○ Video recognition
What’s next?Our current R&D
@pulpix
PARIS NEW-YORK
124 rue d’Aboukir75002 Paris, France
584 Broadway New York10012 NY, USA
Pulpix
Pulpix Inc.
+33 (0)6 66 15 02 42 +1 (415) 996 4453
www.pulpix.com
Pulpix is [email protected]
