Research On Knowledge Enhanced Explainable Recommendation

Recommender system can find user’s potential interests from massive amounts of data,thereby effectively alleviating the“information overload”problem.It has been widely studied and applied in academia and industry.Accuracy and explainability are two important factors to measure the performance of the recommendation.Explainable recommendation can provide reasons for user to help them make better decisions,which becomes a hotspot.Knowledge graph contains abundant facts and semantic relations,which is promising to improve the accuracy and explainability of the recommendation.However,it is usually a large graph with complex heterogeneous and high order connections.How to use knowledge graph to provide accurate and explainable recommendation results has become a challenging problem.This thesis focuses on the explainable recommendation based on the knowledge graph,and studies the problem of: 1)how to effectively integrate graph neural network with heterogeneous knowledge graph; 2)the dynamic evolution modeling of user’s preference and the explainability in knowledge-enhanced sequential recommendation; 3)how to jointly optimize the accuracy and explainability in a unify recommendation framework.To track these challenges,this thesis proposed four solutions:(1)Aiming at the integration of graph neural network model with heterogeneous knowledge graph in explainable recommendation,this thesis proposes a Hierarchical Attention Graph Convolutional Network Incorporating Knowledge Graph for Explainable Recommendation,HAGERec.In this model,the graph convolutional network is used to model the local adjacency structure of entities,and a bidirectional information propagation strategy is designed to aggregate the neighbor information and capture the higher-order semantic relations between entities.At the same time,a hierarchical attention mechanism is designed to adaptively capture and adjust the interaction information of each user-item pair.Finally,the model provides explainable recommendation results through weighted knowledge-aware paths.Experimental results show that this model not only achieves better recommendation accuracy in CTR prediction task and Top-K recommendation task,but also can provide knowledge-aware explainable recommendation results.(2)Aiming at the dynamic evolution modeling of user’s preference in the knowledgeenhanced explainable sequential recommendation,this thesis proposes a knowledge-enhanced hierarchical self-attention model for sequential recommendation,named HSRec.It can model the dynamic evolution of user’s preferences by combining the knowledge graph and sequence information.Specifically,HSRec models the fine-grained preferences through designing inherent interest module and latent interest module.Meanwhile,knowledge paths can be generated for each user-item pair by combining the knowledge graph,and a hierarchical self-attention mechanism is proposed to further explain the dynamic evolution of user’s preferences.Finally,experimental results show that the proposed model has stronger recommendation performance.(3)Based on the HSRec,we further propose a knowledge-enhanced general framework for explainable sequential recommendation,named GFE.Unlike traditional explainable recommendation models,GFE has stronger generality,which can be integrated with other pure sequence recommendation models and endowed them with explainability.In order to verify its generality,two GFE-based models are proposed,called GFE-SASRec and GFE-Ti SASRec.Specifically,GFE-SASRec is obtained by integrating the GFE and SASRec,GFE-Ti SASRec is obtained by integrating the GFE and Ti SASRec.Finally,experiments on real datasets show that the model not only achieves better results on metrics such as HR and NDCG,but also can explain the sequential evolution of user’s preferences from the micro and macro levels.(4)In order to jointly optimize the accuracy and explainability in a unify recommendation framework,this thesis proposes a knowledge-enhanced explainable text generation model for recommendation,named KGERec.It can accurately predict user’s preferences and generate high-quality text by integrating aspect-level information and knowledge graph.KGERec respectively model recommendation accuracy and explainability as rating prediction task and text generation task.These two tasks are integrated into a multi-task learning framework.Besides,in order to evaluate the explainability of the recommendation,this model designs a Graph-toText metric to measure the quality of the generated text,which can further expand the quantitative metric of the recommendation explainability evaluation.Experimental results show that KGERec not only achieves better results on metrics such as BLEU and ROUGE,but also can generate informative and topic-relevant reviews.