Recommendation System with Reasoning Path Based on DQN and Knowledge Graph

Abstract

Recommendation system is a popular research field. In the age of information explosion, a reliable recommendation system is necessary for users. There are a certain number of approaches to do recommendation work. Reinforcement learning is one of the methods used in recommendation system. In this paper, we use reinforcement learning to recommend items to target users, and achieved a rather good result. To give a better user experience, we have added explanations for recommended items. The explanation is realized by Knowledge Graph. We use TransE to embed target users and items, and it helps manage the information of users and items. Our method KGDQN combines Knowledge Graph and reinforcement learning, which can decide the proper recommendation items, and find the reasoning paths from target users to recommended items. Redundant edges are pruned and the DQN model renders a reward function which gives back the result of recommended items, and the explanation paths of the recommendation. Experiments are conducted on Amazon datasets which show the superior performance of KGDQN.