Reinforcement Learning in Recommender Systems

This thesis presents enhancements for the DRR (Deep Reinforcement Learning based Recommendation framework) system (Liu et al., 2019). The original implementation was taken from the open repository RePlay (SB-AI-Lab, 2019). Special attention is paid to increasing the speed and efficiency of the algorithm. We optimized DRR, significantly reducing the training time. Inspired by the article Tiapkin et al. (2022), we developed and implemented a new method, Bayes-DDPG. This approach uses Bayesian exploration with a multi-headed critic and Gaussian noise. Experiments on the MovieLens dataset using NDCG, HitRate, and Coverage metrics showed a significant improvement in the quality metrics of the optimized DRR system compared to the original implementation. The algorithm enhancements are published in open access, allowing other researchers and practitioners to use and develop our approach.