Diversity Preserving Explainable Reinforcement Knowledge Graph Recommendation System

The recommendation system based on the reinforcement knowledge graph makes use of the rich knowledge information in the knowledge graph and the powerful exploration ability of reinforcement learning to effectively alleviate the data sparsity problem.However,current recommendation algorithms generally focus more on the accuracy of recommendations and less on the diversity of recommendations,resulting in the final recommendation results being too homogeneous and homogeneous.Therefore,it becomes a challenge to maintain recommendation diversity while ensuring the accuracy of recommendation results.The key issue in improving recommendation diversity is to control the degree of diversity based on users’ personalized needs and to quantify the diversity of recommendation items.It is difficult to model users’ behaviors globally with single-layer knowledge graph-based recommendation reasoning,and further research is needed to infer users’ real preferences and further improve the accuracy of recommendations.Based on the above problems,the main contents of this paper are as follows:(1)A personalized diversity recommendation algorithm DRA-UIN based on reinforcement learning and knowledge graph was proposed to control the degree of diversity by predicting users’ diversity needs.Firstly,using the user as the starting point,the Agent is used to wander through the knowledge graph in a supervised pathfinding manner to evaluate and score the user’s recommendation diversity needs;then the Actor-Critic method is used to find personalized recommendation items based on the user’s diversity needs score,and the policy function Actor is used to guide the Agent to perform path reasoning in the knowledge graph;finally,a new Finally,a new diversity metric is proposed to quantify and reward the diversity of recommended items,which helps train the policy network to recommend items more accurately.The experimental evaluation on several large-scale real datasets shows that DRA-UIN effectively improves the diversity and accuracy of recommendation results.(2)The reinforcement knowledge-guided multilevel reasoning method MRPD for maintaining diversity is proposed to further improve the accuracy of recommendations by exploring the real interests of users through top-down multilevel recommendation reasoning in the instance view and ontology view of the knowledge graph,which solves the problem that single-level knowledge graphs cannot globally model user behavior in recommendation reasoning.First,we wander through the multi-level knowledge graph to obtain deeper user diversity requirements;then,we use the knowledge in the high-level knowledge base to guide the underlying reasoning;finally,we quantify the diversity of recommendation items by judging the diversity of reasoning paths and the discrete degree of recommendation item parent ontology,and design a reward function to help push up the diversity of recommendations.Experiments show that this method outperforms several baseline methods in terms of recommendation accuracy.