Research On Knowledge-enhanced Recommendation Method Via Graph Attention Networks

As a mature information filtering technology,recommender systems effectively solve the information overload issue and enhance the user experience by modeling users’ latent interests and preferences.Knowledge Graphs,which consist of real-world connected rich entities,play a vital role in alleviating the inherent sparsity and coldstart issues of recommendation systems.A dominant technology trend is to design endto-end Graph Neural Networks-based Knowledge(Graph)-enhanced recommendation models.However,current Graph Neural Networks-based knowledge-enhanced recommendation approaches focus on extracting high-order neighbor attributes of knowledge graphs,ignoring three key aspects to improve recommendation quality:Firstly,the extraction of finer-grained feature interaction signals in knowledgeenhanced tasks.Secondly,the representation of complex intent behind user-item interactions.Thirdly,the mining of high-quality self-supervised signals in graph recommendation models.The above issues hinder the performance of recommender systems to a certain extent.To this end,this thesis presents an exhaustive summary and analysis work of existing knowledge-enhanced recommendation work,and the following progress has been made so far based on existing research,covering:1.A knowledge-enhanced recommendation method that integrates feature interaction and intention-aware attention networks is proposed.Firstly,a knowledgeenhanced backbone network based on Graph Attention ne Tworks is used to generate user/item prototype representations.Secondly,a dual-grained convolutional neural network is used to extract finer-grained feature interaction signals by performing vertical and horizontal convolution,respectively,to enhance item-side representation learning capability.Finally,based on the idea of disentangled representation learning,a two-level attention mechanism is employed to perform latent intent behind user-item interaction to enhance the user-side representation learning capability and thus further improve the recommendation performance.2.A knowledge-enhanced recommendation method based on co-contrastive learning of dual-view is proposed,which integrates two semantically complementary but different views implied by the Collaborative Knowledge Graph topology.Firstly,for the subjective interaction view,the prototype user/item representation is modeled using light-weight Graph Convolutional Networks and a divergence regularization term is imposed for further optimization.Secondly,for the Knowledge Graph objective structural view,a knowledge-enhanced attention network is utilized to extract highorder features.Next,the dual-grained Convolutional Neural Networks are used to perform vertical and horizontal convolution methods to capture finer-grained feature interaction signals.Finally,high-quality co-contrastive learning is performed between the dual-view,thus improving the node representation learning capability and enabling mutual supervision and enhancement.Extensive experiments on three real-world Knowledge Graph datasets show that the two knowledge-enhanced recommendation methods proposed in this thesis outperform the baseline models compared in this thesis and clarify the effectiveness and rationality of the proposed models.Numerous hyperparameter experiments and ablation analyzes unveil the mystery of the models and the complex mechanism behind them.