Research On Recommendation Models Based On Knowledge Graph

In recent years,recommender systems have gained increasing attention from both industry and academia,as they play a crucial role in helping users navigate through vast amounts of information.One promising approach for improving recommendation accuracy is to integrate auxiliary information into the models.Knowledge graphs have emerged as a comprehensive and popular option for this purpose.However,existing knowledge graph-based recommendation models still face some limitations that require further research.On the one hand,current models do not leverage ”entity” information to infer user intents.The weighted sum method is used to calculate the feature embeddings of user intents,which results in insufficient discrimination of user intents across different users.On the other hand,the current knowledge graph-based recommendation models don’t fully consider the popularity bias and the impact of item popularity on user behavior in recommendation scenarios.To address these issues,this thesis proposes two novel knowledge graph-based recommendation models.The main work of this thesis is as follows:1.The relevant literature on knowledge graph-based recommendation models,graph neural networks,and popularity bias is reviewed,and the advantages and disadvantages of existing models are analyzed.2.An entity-driven user intent inference knowledge graph-based recommendation model(EKIN)is proposed.The existing models only use the ”relation” information in the knowledge graph,but ignore the ”entity” information,and have the problem of insufficient discrimination between different user intent embeddings.To address these issues,EKIN uses a graph to model user intents and builds an entity-driven user intent graph(EUIG)that represents an user intent with entities in the knowledge graph as the core.EUIG utilizes both entity and relation information and introduces statistical information.A frequencybased attention computation method is proposed to learn on the EUIG,which combines the information of embeddings and statistic.EKIN uses graph neural networks to learn in the knowledge graph and makes personalized recommendations based on each user’s intent,thus improving recommendation accuracy.3.A gated alleviating popularity bias knowledge graph-based recommendation model(GABKIN)is proposed to address the issue that the existing models do not fully consider the popularity bias and the influence of the popularity of items on user behavior in recommendation scenarios.Building on the EKIN model,GABKIN enhances the overall structure and explains user interaction behavior through causal graphs.GABKIN decomposes user interaction behavior into preference-based and item popularity-based interactions through causal inference.The gating mechanism is designed to distinguish the influence of user preference and item popularity on user behavior.GABKIN alleviates popularity bias and improves the accuracy of the recommendation model.4.EKIN and GABKIN are experimented and evaluated on various widely used realworld datasets.The experimental results demonstrate that the proposed models improve recommendation accuracy.