Commodity recommendation system based on deep reinforcement learning

• 1、School of Computer, Beijing University of Posts and Telecommunications, Beijing,10086

Abstract：Current commodity recommendation systems are often based on the fact that user data has been fully obtained, and their behavior characteristics will not change in a long time. However, users and recommendation systems often have continuous and close interactions, which can better reveal current user behavior characteristics and provide more basis for recommendation systems to make accurate recommendations. In response to this problem, this article has mainly done two aspects of work. First, this paper designs and implements a multi-element approaching state mechanism and an action grouping mechanism. The multi-element approaching state mechanism makes it possible to obtain more credentials when obtaining the reward value of the similar state, and to obtain more similar state elements. The action grouping mechanism is to collect the same action as a group, reducing the amount of calculation for each state. Second, this article studies the expansion and optimization of the product recommendation system. The system can support multiple updates and implementations of deep reinforcement learning recommendation algorithm modules. It is designed and implemented to support user login, product recommendation, user product operations, and user operations. Recorded commodity recommendation system.

Keywords： Deep reinforcement learning, multi-element approaching state, Action grouping