Research On Graph Neural Network-based Recommendation Algorithm With Entity-level User Preference

Knowledge graph has been introduced to the recommendation field and developed knowledge graph-based recommendation algorithms because of its ability to enhance the semantic connections between data through rich entities and relations,thus improving the accuracy and explainability of the recommendation results.Using the structure information and semantic information of knowledge graph through graph neural network to improve the recommendation results is an important one of them.Although current approaches have made good progress,they still face the problems of modeling user preference roughly and the existence of noise in the data.To be specific: existing methods either do not model the fine-grained preference behind user-item interaction,or model preference only at the relation level of the knowledge graph,ignoring the user preference at the attribute entity level;the improvement of the recommendation results relies on high-quality knowledge graph and interaction data,however,the limitations of the construction methods lead to noisy entities in the knowledge graph,and the inaccuracy of implicit feedback makes the data contain noisy interactions such as misclicks.Such problems are not conducive to the learning of user and item representations and restrict the recommendation performance.Therefore,it is of great theoretical significance and application value to study how to better model user preference and alleviate the noise problems in the data to improve the algorithm accuracy and user satisfaction.To address the above problems,this thesis proposes two recommendation algorithms from the data-driven and the priori knowledge-driven perspectives to model user preference to the attribute entity level by exploiting the rich connectivity information in the graph structure and alleviate the noise problems in the data using user preference.The main contributions of this thesis are as follows.(1)To address the problems of lack of modeling fine-grained preference behind user interaction and noisy entities in knowledge graph in previous method,this thesis proposes a contrastive learning recommendation algorithm based on the attention mechanism to model entity-level user preference from the data-driven perspective.First,the algorithm extracts key attribute entities of interest to user through graph structure and retains such entities in the construction of contrastive graphs of knowledge graph to prevent the loss of key entities from making the semantics of item dominated by noisy entities;then,the algorithm learns the representation of entity-level user preference from the data through the attention mechanism and incorporates it into the construction of contrastive graphs of interaction graph to reduce the influence of noisy entities on semantics of user;finally,the algorithm optimizes the user and item representations using contrastive learning to further improve the robustness of the model against noisy entities.(2)To address the problems of the previous method that only models the preference behind user interaction at the relation level and noisy interactions in implicit feedback,this thesis proposes a recommendation algorithm that models entity-level user preference based on the priori weights of the number of connections from the priori knowledge-driven perspective.First,by introducing the number of connections between attribute entities and items interacted by user and fitting a priori weights for them to model user’s preference for entities;then,using the priori weights to guide the learning of user entity-level preference representation;finally,using the user preference to score the items interacted by user,thus reducing the weight of noisy interactions in user representation learning and improving the accuracy of user representation.(3)Extensive experiments are conducted on several datasets in this thesis,and the results demonstrate that modeling user preference to the attribute entity level can not only capture user preference more granularly and improve recommendation performance,but also alleviate the problems of noisy entities and noisy interactions in the data effectively and improve the robustness of the models to the noisy data.Thus,the validity of the proposed methods is proved.