Reinforcement Learning Recommendation System Based On Heterogeneous Information Network

By continuously interacting with users,reinforcement learning-based recommendation systems are able to adapt to changes in user preferences and take into account the potential future benefits of a product.However,there are still some challenges,such as the variability,sparsity and complexity of user interaction data.Incorporating heterogeneous information networks into reinforcement learning methods can capture the topology and potential connections between users and products,alleviate data sparsity,and consider the long-term value of recommended products.Inspired by this,this paper presents three works around the problem of how to effectively utilize auxiliary information in reinforcement learningbased recommender systems as follows:(1)A knowledge-guided adaptive sequence-based reinforcement learning model is proposed,which adaptively truncates drift sequences of different lengths through the feature information of the knowledge graph and the user-item interaction sequences,using which can enhance the exploration capability of the recommendation system and capture the preference drift of users in a timely manner.In addition,a new composite reward function is designed with the feature information of the knowledge graph,including discount sequence reward and knowledge graph reward,which can alleviate the problems of reward sparsity and slow convergence.(2)A multi-layer attentional social recommendation framework based on deep reinforcement learning is proposed to learn the dynamic preferences of users through deep reinforcement learning methods and end-to-end approach.A structured representation of users and items is constructed using subgraphs as an adaptation to the interactive recommendation framework.When there is a lack of user interaction data in the user-item bipartite graph,the personalized preferences of the target users are indirectly analyzed by using the information transfer between socially trusted neighbors in a heterogeneous information network.In addition,the state representation module is designed with a multilayer attention mechanism to mitigate the influence of noisy nodes within the subgraph and improve the personalized representation of states.(3)A social multi-intelligence reinforcement learning framework is proposed.Firstly the model enhances the recommendation performance by constructing two intelligences.Secondly,corresponding state representation modules are designed for different intelligences for learning state representations from social networks and user rating matrices,respectively.Finally,to efficiently utilize the neighbor nodes in the social network,advanced social neighbor features are aggregated using trust inference in addition to firstorder social neighbor features to alleviate the problems of data sparsity and cold start.