Research Of Knowledge Graph Based Explainable Recommendation

In the era of big data,recommender system can find out the potential demand by analyzing users' history behavior data,and then recommend users information to solve the problem of information overload.However,the traditional recommendation algorithm has the problem of data sparsity.Using different types of auxiliary information to supplement interactive data can solve the above problem and improve the recommendation performance.Integrating knowledge graph as auxiliary information into recommendation algorithm can not only introduce more semantic relations to improve the recommendation performance,but also make it possible to explain recommendations.Path-based recommendation is main method of knowledge graph based recommendation.In existing work,path is generally regarded as an entity sequence,and entity representation is input into recurrent neural network to learn path representation in the order from beginning to end of the sequence.Therefore,the quality of entity representation and path representation influences recommendation performance.So far,existing recurrent neural network based path representation learning methods have achieved a series of research results,but there is still room for improvement.In particular,existing entity representation learning methods only focus on entities and relationships in triples,and do not consider the information of the neighboring entities,which affects entity representation learning.In addition,existing path representation learning methods are often limited by the structure of recurrent neural networks,which representation can only learn the sequence dependencies among entities within a path,and also lack the consideration of the direct dependencies among non-adjacent entities within the path,leading insufficient path representation learning.In order to solve the above problems,this thesis proposes a knowledge graph based explainable recommendation algorithm KTRec,leveraging the path representation learning with self-attention mechanism.The contributions of this thesis are as following:(1)this thesis designs a pre-training method,which combines graph convolution network and matrix factorization,to learn entity representation.Thus,the learned eneity representation combines the auxiliary information,implicit feedback and explicit feedback in the user history preference information.(2)this thesis proposes a path representation learning method with the self-attention mechanism.This method uses the self-attention mechanism of Transformer model to calculate the attention distribution between each entity representation within a path,and generate the path representation according to the distribution,which fully considers the dependence among the entities.(3)this thesis designs an optimal path representation method based on dual attention,and further optimizes the path representation by intergrating user attention module and item attention module.Experimental results show that the proposed KTRec algorithm achieves higher recommendation accuracy than existing works.In addition,this thesis also designs a knowledge graph based explainable recommendation prototype system to provide users with KTRec based personalized explainable recommendation service.The prototype system includes Web presentation module,recommendation model module and data module.In particular,the recommendation model module includes model training,recall phase and sorting phase.The data module includes data collection and data storage.After the design and implementation of each module,the prototype system meets the needs of both functional and non-functional requirements in the requirement analysis phase.