Explainable Recommendation Method Based On Knowledge Graph

With the rapid development of the Internet,information is showing explosive growth.However,it is difficult for users to obtain and use effective information from the huge amount of data.This phenomenon is called information overload.Recommendation system is one of the effective methods to solve the problem of information overload.At present,the recommendation system has been applied in various web applications because it can recommend the content based on the information of users and items.Most recommendation systems generally use collaborative filtering method which has a weakness in interpretability.Because knowledge graph has rich semantic features and strong interpretability,it can be introduced into the recommendation algorithm to solve the problems above.However,there still remains two problems in the current recommendation methods based on knowledge graph.Firstly,the recommendation method based on knowledge graph path uses random method to extract path,which results in too much noise.Secondly,the recommendation method based on knowledge graph is difficult to extract the implicit relationship features between items outside the knowledge graph.To solve these problems,this thesis proposes an explainable recommendation method based on knowledge graph.First of all,this method builds a domain knowledge graph which is suitable for recommendation and uses reinforcement learning to extract the paths between users and items in the knowledge graph.Then,it extracts the explicit relationship features between users and items in the knowledge graph by modeling the path and extracts the implicit relationship features between items outside the knowledge graph by using the user's historical interactions.Finally,this thesis employs UCL to label knowledge graph and develops a prototype system.The details are as follows:(1)In order to extract the path between users and items which is more suitable for recommendation,this thesis proposes a Path Extraction of knowledge graph based on Reinforcement Learning(PERL),which uses the state and action set as the input of the ActorCritic network and then gets the score of each action in the current state.The knowledge graph representation task and recommendation task are used in the training to speed up the convergence of the model.Finally,the strategy path selection algorithm is designed to extract the paths between users and items.(2)Considering the current methods based on knowledge graph can only capture the explicit relationship features in knowledge graph and difficult to discover the implicit relationship features outside the knowledge graph,this thesis proposes an Explainable over Knowledge-aware Path Network for recommendation(EKPN).The method consists of the Attention Knowledge-aware Path Module(AKPM)and the Implicit Feature Extraction Module(IFEM).AKPM can capture explicit relationship features between items in the knowledge graph by paths which are extracted by PERL from users to candidate items for recommendation;IFEM explores implicit relationship features between items outside the knowledge graph by utilizing users' historical interactions.At the end,the method uses a fusion mechanism to combine the two modules to achieve higher performance.(3)A prototype system for recommendation system based on knowledge graph is implemented,which could show the recommendation results and reasons.Two real-world datasets are used for comparison and ablation experiments in order to prove the effectiveness of the method proposed in this thesis.Comparison experiments show that the PERL and EKPN methods proposed in this thesis have better performance than the baseline model.The ablation experiment not only proves the rationality of each module in the model,but also proves that the path between users and items extracted by the PERL can better mine users' interests and improve the recommendation performance.Finally,it is proved that the proposed methods are interpretable by case analysis.