Research And Analysis Of Explainable Recommendation Algorithm Based On Graph Networks

With the development of the mobile internet and the information society,people are in a sea of information.Considering the information explosion,the recommendation system which provide viewers with required information was proposed.It makes recommendation algorithms popular in current research.The recommended field is not limited to item recommendation,but is applied to various fields,such as judicial recommendation(recommend relevant laws based on a user's description),news recommendation,service recommendation,and so on.Users can avoid information explosion and improve the efficiency of information services in various fields with the recommendation algorithm.The traditional recommendation algorithm has three major shortcomings.Firstly,the structural construction of features is insufficient,including the ig-norance of text relations,ignorance of specific scenes and so on.Secondly,previous recommendation algorithms designed multi-layer network with sim-ple fusion of features which lose important information easily during training.Thirdly,the existing recommendation algorithms have insufficient explainability due to the black box characteristics.With the above problems,three recom-mendation algorithms have been proposed in this paper.First of all,we propose a recommendation algorithm(HMN for short)based on a hierarchical matching mechanism.HMN leverages the semantic information of the candidates to be recommended,and first divides them into a hierarchical structure based on the attributes of the candidates to be rec-ommended.Then we propose the matching mechanism to turn the task into a matching problem.HMN leverages the candidate semantics and structure neglected by previous work and utilizes the matching mechanism to make ex-planatory recommendations.Secondly,we propose a recommendation algorithm based on graph neural network(GLAM for short).Different from the homogeneous graph constructed in the previous work,GLAM designed a heterogeneous graph based on the re-lations.Then we design a graph convolution and the matching mechanism for information extraction and explanatory recommendation.Finally,the experi-ments on two real datasets show the effectiveness of GLAM.Finally,we propose a knowledge-guided reinforcement learning recom-mendation algorithm(KERL for short).Considering the limitations of neural networks models,KERL explores the feasibility of applying reinforcement learning to the sequential recommendation task.By defining the task as a Markov decision process,KERL captures the long-term interest of item se-quence for sequential recommendation.KERL constructs the knowledge graph to guide the exploration process of reinforcement learning,so KERL is not limited to recommendation,but can also predict future preferences.Experimental results on real-world datasets demonstrate that our proposed algorithms have achieved significant improvements over all of the baseline methods.Two papers written with above innovations have been accepted and published by CCF-A international conferences.One paper has been accepted and published by the top domestic information retrieval conference.