Self-learning Recommendation System Using Reinforcement Learning

Abstract

Recommendation system are widely used in e-commerce that is a part of e-business. It helps users locate information or products that they would like to make offers. In this paper, we purpose a new web recommendation system based on reinforcement learning, which is different from another system using Q-learning method. By using ε-greedy policy combined with SARSA prediction method, another powerful method of reinforcement learning is obtained. The system gives customer more chance to explore new pages or new products which are not popular which may match with their interests. The system composes of two models. First, a global model, the model for all customers to discover behavior of system. We can know another users direction or trend by global model. Second, a local model, which uses to keep records of user browsing history and makes offer from each customers. We report experimental studies that show the click rate of recommendation list.