Reinforcement Learning for Long-term Reward Optimization in Recommender Systems

Recommender systems help users to orient in the vast space of goods, services, and events. A user interacts with the recommender engine in a sequence of exchanges of recommendations and user feedback. The idea that previous interaction influence the later ones and the importance of the sequence of interactions can be modeled using Markov decision processes and solved by reinforcement learning. Several recent articles applying reinforcement learning to recommender systems have proved the viability of this direction. But it is still difficult to compare different approaches. We propose an environment with a unified interface that will permit to compare different modelization of recommender process and different algorithms on the same underlying sequential data. We also performed the extensive parameter study for deep deterministic policy gradient methods on the well-known MovieLens dataset.