Research On Item Collaborative Filtering Recommendation Algorithm Based On Graph Neural Network

Collaborative filtering is the most widely used recommendation algorithm in the recommendation system.Item-based collaborative filtering has strong interpretability,realtime performance and accuracy.Given that most data in recommendation can be organized by graph,and graph neural network has shown its great power in aggregating low and high order neighborhood information,collaborative filtering recommendation algorithm based on graph neural network has drawn widespread attention from researchers.However,there is still further space for improvement and enhancement in the existing work.Therefore,this paper analyzes the existing research work and studies the research on item collaborative filtering recommendation algorithm based on graph neural network from three aspects:graph structure modeling,neighborhood information noise and data sparsity.Firstly,the homogeneous graph neighborhood aggregation recommendation method for items is proposed.This method models the item-item homogeneous graph from the perspective of graph structure modeling,enhancing the closeness between nodes,and utilizes graph convolutional neural networks to aggregate collaborative information among item neighborhoods.Additionally,a residual connection factorization module is designed from the perspective of reducing neighborhood information noise to alleviate the potential impact of noise on model performance,thereby improving the accuracy of the model’s recommendations.Secondly,Aiming at the problem of data sparsity,this paper proposes an item information enhancement recommendation method based on self-supervised learning.This method uses self-supervised contrastive learning strategies to construct positive and negative sample auxiliary supervision signals and combines supervised and self-supervised tasks using a multi-task learning strategy to alleviate the problem of data sparsity.In order to enhance the performance of the model,an attention-based layer information fusion module is designed to personalize the information fusion weights of each convolutional layer,enabling the model to better learn item embedding representations and improve the model recommendation performance.Finally,this paper conducts experiments on three datasets Music,Last-FM and Yelp using the Recall and NDCG evaluation metrics and compares the results with baseline methods,including FM,FISM,NAIS,GCMC and Light GCN.The key modules and important experimental parameters are studied and explored.The effectiveness of the model proposed in this paper in improving the recommendation effectiveness is analyzed and verified.