Dr. Michael Roizner
In many Yandex services, users are choosing and consuming various types of content such as music, movies, goods, videos, news, etc. Recommender systems are known to significantly improve user experience and the key quality metrics of such services. My talk describes our unified approach to recommendation. We present an extensible model that supports different domains, content features, external data both for users and items, and real-time learning. It uses both collaborative filtering and content-based algorithms, combining them via Yandex’s well-known algorithm MatrixNet. Our method can be adopted to different metrics appropriate in various recommendation tasks.