Deep Reinforcement Learning based Recommend System using stratified sampling

Abstract

Typically personalized movie recommendation algorithms often adopt a static view of recommend process and only take current rewards into consideration. Thus, they are hard to adapt to the dynamic change of users and items. In this paper, we propose a movie recommend system based on deep reinforcement learning to better accommodate the dynamic property when users' distribution or interest changes. Firstly, we adopt nature DQN algorithm to set up baseline. Second, under the framework of nature DQN, we use Double DQN to solve overestimation and indeed reduce error. Besides, we use stratified sampling rather than random sampling to accelerate convergence. Finally, by testing on Movielens dataset, the experimental results shows that our algorithm is superior to traditional algorithms, and also comparable to the latest algorithms.