recommendations at senscritique. mixing social and machine learning
TRANSCRIPT
Recommendations at SensCritique Mixing social and machine learning
By Xavier RAMPINO, Mars 2016
What is SensCritique ?
What is SensCritique ?
What is SensCritique ?
Entertainment database
What is SensCritique ?
Movies, TV Shows,
Video Games, Books,
Comics, Music
Entertainment database
What is SensCritique ?
Movies, TV Shows,
Video Games, Books,
Comics, Music
Entertainment database Virtual library
What is SensCritique ?
Movies, TV Shows,
Video Games, Books,
Comics, Music
Entertainment database Virtual library
Rate, List,
Review, add to a wish list
What is SensCritique ?
Movies, TV Shows,
Video Games, Books,
Comics, Music
Entertainment database Virtual library Social network
Rate, List,
Review, add to a wish list
What is SensCritique ?
Movies, TV Shows,
Video Games, Books,
Comics, Music
Entertainment database Virtual library Social network
Rate, List,
Review, add to a wish list
Users can : Follow each other,
like activities, recommend products
What is SensCritique ?
Database + Library + Network
What is SensCritique ?
Database + Library + Network
What is SensCritique ?
=
Database + Library + Network
What is SensCritique ?
=
What is SensCritique ?
Database + Library + Network
What is SensCritique ?
Database + Library + Network
What is SensCritique ?
Database + Library + Network
Discovery Engine
What is SensCritique ?
Discovery tool
Database + Library + Network
Discovery Engine
What is SensCritique ?
Discovery tool
Discovery tool
Database + Library + Network
Discovery Engine
What is SensCritique ?
Discovery tool
Discovery tool
Discovery tool
Database + Library + Network
Discovery Engine
What is SensCritique ?
14M
600K
55M
1,5M
products
registered users
ratings
lists
How we make it work
How we make it work
How we make it work
• Multimedia Database
Entertainment database
How we make it work
• Multimedia Database
• Extensive Cinema, TV, and VOD showtimes
Entertainment database
How we make it work
• Multimedia Database
• Extensive Cinema, TV, and VOD showtimes
• Graph Database
Entertainment database
How we make it work
Key factors for discovery :
• Serendipity • Familiar cues (actors, franchises) • Familiar media (TV, cinema)
Entertainment database
How we make it work
Virtual Library
We have very enthusiastic members
How we make it work
Virtual Library
Our past recommendation engine stack
How we make it work
Virtual Library
Our past recommendation engine stack
Ratings
Wish
CSV dump LensKit
MySql
How we make it work
Virtual Library
Our past recommendation engine stack
Very extensive operation, once a day
Ratings
Wish
CSV dump LensKit
MySql
How we make it work
Virtual Library
Our current recommendation engine stack
How we make it work
Virtual Library
Our current recommendation engine stack
Nginx Logs
Web hooks
prediction.io event server Application
How we make it work
Virtual Library
Our current recommendation engine stack
Nginx Logs
Web hooks
prediction.io event server Application
• Live Update • Take product views into account • Live boosting during querying (via Elasticsearch)
How we make it work
Social Network
We use social recommendations at two levels
How we make it work
Social Network
We use social recommendations at two levels
User-User recommendation Live Feed
Direct
How we make it work
Social Network
We use social recommendations at two levels
User-User recommendation Live Feed
Direct
News filteringTops
Aggregated
How we make it work
Once one has chosen people to follow, he gets live update on their entertainment activity
Social NetworkDirect
How we make it work
He may also be given direct recommendation
Social NetworkDirect
How we make it work
To allow one member to find similar members, we made a user-user recommendation engine using Mahout :
Social NetworkDirect
Online recommendation system, webservice served over Glassfish Event updates are sent on RabbitMQ, then dispatched to a Java Consumer
How we make it work
One member can notice than a subgroup of member he follows did an activity on a specific product
Social NetworkAggregated
How we make it work
Community choices, given an editorial theme Their own activity (we’ll show them only polls for which
they already rated some of the result products)
Social NetworkAggregated
We have also polls, in which all members can participate. Member then get recommendations fuelled by two forces :
How we make it work
Social NetworkAggregated
Conclusion
From its origins, SensCritique is headed towards product discovery, as a whole though 3 ways :
• Database : We tend to have a comprehensive and ambitious graph-oriented database
• Library : We are a tool for the user to keep a memory of his entertainment history
• Social : We foster social activity between our members, and we are proud of our community
We strongly believe that this is thanks to these three factors that we can offer both a tailored and a global recommendation offer to our audience.
Questions ?
Question No question