Abstract: In order to solve the cold start problem caused by traditional recommendation algorithms when they face the recommendation task with sparse data,this paper introduces the knowledge graph into the recommendation algorithm,combing a new neighbor perception attention mechanism to replace the traditional graph attention mechanism to mine the higher-order connected information between entities,and proposes a recommendation model KGNPAN based on the knowledge graph and neighbor perce-ption attention mechanism.Thanks to the knowledge graph,recommendations can be accurate,diverse and interpretable.This model can effectively alleviate issues of data sparsity and cold start.Firstly,this model utilizes the graph embedding method RotatE based on self adversarial negative sampling to expand the semantic information of the original item and user representations,mapping entity and relationship vectors into low dimensional embedding vectors.Secondly,based on the different types of collaborative neighbors,neighbor perception attention mechanisms are applied to aggregate neighbor node information,enrich the semantics of target nodes,and recursively mine high-order connected information in convolutional form.Finally,the inner product operation is applied to calculate the interaction probability between the user and the project vector,and the recommendation result is obtained.Experiments are conducted on two common benchmark datasets,Amazon-book and Last-FM,and compared with six benchmark models,namely CKE,BPRMF,RippleNet,KGAT,KGCN,and CAKN,KGNPAN.The results show that KGNPAN improves the recall rate by 1.30% and 1.37%,and normalized discounted cumulative gain(NDCG) increases by 1.26% and 1.14%,respectively,compared with CAKN model,which has the best performance in the benchmark modes,verifying the effectiveness and interpretability of the model.

Key words: Recommended algorithm, Neighbor perception attention mechanism, Knowledge graph, Graph neural network, Cold start