Iterative Traffic Control Algorithm to Solve the Article Cold-Start Problem at Yandex.Zen

This work is devoted to the article cold start problem within the Yandex.Zen recommendation system. New publications are supplied with exploration shows, helping the recommendation system to gain additional collaborative information to solve the problem. This work is divided into several parts. The first part examines the impact of exploration shows on the Yandex.Zen recommendation system, namely on the ranking of new publications and on channel coverage growth. As a result, it is concluded that exploration shows are particularly important for new publications written by authors with a small number of readers. The second part describes a traffic control algorithm that efficiently distributes a limited number of exploration shows between multiple new publications. The considered algorithm has several advantages over the previous scheme for managing exploration shows, and is implemented as a component of the Yandex.Zen recommendation system. The paper also describes the implementation of an industrial environment simulator used for the algorithm’s parameters grid search. An AB-experiment conducted as part of this work shows that the presented algorithm allows doubling the exploration shows clickability. An example is also given which demonstrates that the algorithm can be successfully applied to solve other traffic control problems. In addition, this work describes in detail The Yandex.Zen recommendation system and its architecture as well as practical application of ALS decompositions in collaborative filtering. The literature survey on the problem of article cold start problem is provided.