Multi-layer Attention Social Recommendation System Based on Deep Reinforcement Learning

Abstract

The recommendation system based on deep reinforcement learning recommends interesting content to users through the interaction of recommendation agents and usersThis is to inform you that corresponding author has been identified as per the information available in the Copyright form.. However, most recommender systems based on deep reinforcement learning often face two limitations: (1) sparse user feedback data that makes the recommendation agent unable to accurately capture the user’s dynamic preferences; and (2) users and items are isolated from each other due to the limitation of unstructured representation. To address this situation, this paper proposes a multi-layer attention social recommendation method based on deep reinforcement learning by fusing social network and user-item bipartite graph to form a heterogeneous information network. This method uses subgraphs in the heterogeneous information network to structurally represent users and items through a variant of the graph attention network. By doing so, users and items can perceive neighborhood information in the heterogeneous information network and enhance the correlation between nodes. At the same time, it can also avoid the repeated transmission of irrelevant nodes. Additionally, the attention mechanism is adopted in the graph attention network to reduce the influence of noisy nodes in the heterogeneous information network. Corresponding weights are given to the neighborhood information, and nodes with noise are given smaller weights. Then, the external attention mechanism is used to adjust the weight of historical items in the state information, which realizes the selective attention of different users to different items and generates the user’s preference representation. Finally, the deep reinforcement learning method is used to simulate the interaction between the recommendation system and the user. This method adapts to the dynamic changes of user preferences and considers the long-term rewards brought by the recommended items. Experimental results show that this method can alleviate the above two problems and provide users with more accurate recommended items.