RBM-Q Learning for Rate Prediction in Recommendation Systems

Abstract Through the recommendation of personalized information, recommendation systems are essential in improving the user experience on a variety of online websites. Some popular and conventional techniques have not been able to perform well in data volume and sparse data problems. In order to compete with them, this paper presents and assesses two algorithms: RBM-Q-Learning 1 and RBM-Q-Learning 2. By proposing novel mechanisms for state representation and action selection, these algorithms attempt to overcome the shortcomings of conventional methods, utilizing Restricted Boltzmann Machines (RBMs) as a major component. The performance of recommendation algorithms on two datasets from the MovieLens platform—MovieLens 100K and MovieLens 1M—are compared in-depth through an experimental research presented in this research. The experiments have been performed based on MAE, RMSE, HR, ARHR, Diversity and Novelty measurement indicators and the proposed algorithms have found proper stability.