Customer
World’s leading entertainment services broadcasting online TV series, documentaries and feature films . The customer would like to automatically learn user preferences, in order to predict users’ratings for unseen movies and provide most accurate personalized recommendations.
Challenges
- Data records: only <movie ID>, <user ID>, <date>, <rating> … No metadata describing users or movies in any other, more detailed way.
- Algorithms had to track and detect statistical patterns across multiple user-movie interactions, not knowing what particular movies actually contained.
- The details of the customers Cinematch were kept confidential. Designers had to design their algorithms from scratch.
Solution
- Customer specific algorythm was designed from grounds up
- Universal algorithm for recommendation systems was invented. Within this algorithm, a matrix factorization model designed specifically for recommendations, is trained efficiently via gradient descent.