| Graph Convolutional Network(GCN)has become an important technique in the field of Graph Neural Networks(GNN).GCN uses convolutional operations to process graph structured data to learn feature representations of nodes and edges,and has been used not only in academic research but also in practical applications such as social network analysis,recommender systems,and chemical molecular analysis.By using the Knowledge Graph(KG)can provide more comprehensive information to the recommendation system and enable the recommendation system to better understand the relationship between users and items.This improves the accuracy of recommendations and user experience.By using knowledge graphs and graph convolutional neural networks,the efficiency of traditional recommendation algorithms can be effectively improved,specifically,this paper will address the following three key issues.Firstly,to address the problem of how to perform entity embedding aggregation after learning multi-layer neighborhood representation by using the embedding representation of entities enriched with knowledge graph and performing message passing on GCN,this paper proposes the Recommendation Algorithm based on Knowledge Graph and Bidirectional Interaction Graph Convolutional Network(KBGCN),which uses the knowledge-aware attention mechanism to distinguish the neighbor contributions of different views through deep propagation on the knowledge graph,and uses a bidirectional interaction aggregator to successfully aggregate different domains,capturing the item’s knowledge-based higher-order interaction information and improves the recommendation performance.Secondly,to address the problems of heavy GCN network design,high algorithm complexity,difficult model training and overfitting under deep graph convolution,this paper proposes Recommendation Algorithms based on DropEdge and LightGraph Convolutional Networks(DLGCN),adding DropEdge to LightGCN to enrich the input and reduce message passing,and using a weighted average aggregation function to obtain the final representation,which alleviates the oversmoothing problem brought by deep network algorithms,improves the Precision,Recall and NDCG,and enhances the interpretability of the recommendation algorithm.Finally,to address the problem that propagation-based recommendation methods usually ignore the complex relationships between entities in multi-hop ripple sets and do not distinguish between user nodes and other entity nodes,and cannot fully utilize collaborative signals in user-item interactions,this paper proposes Recommendation Algorithms based on Collaborative Knowledge Propagation Graph ATtention Networks,(CKPAT),which can distinguish the importance of different multi-hop ripple sets of users or items,and the ripple set embeddings,user/ item initial entity set embeddings and the original representation of item are weighted by attention aggregator to generate a more fine-grained embedding representation for Top-N recommendations. |
