Research On Recommendation Algorithm Based On Graph Neural Network

At present,the amount of information such as online short video,e-commerce advertising and news is growing rapidly.How to build a more accurate and personalized recommendation system has become the focus of major e-commerce platforms and audio as well as video portals.The core part of the recommendation system is the recommendation algorithm,which directly affects the quality of the recommendation system.Although traditional recommendation algorithms such as collaborative filtering have been developed for many years,it is difficult to solve problems such as sparsity and cold start.On the other hand,due to the development of graph neural network in recent years and its high compatibility with the interactive graph structure of recommendation systems,a variety of recommendation algorithms based on graph neural network come into being.However,these algorithms are more or less faced with the problem of system scalability,and some of them are difficult to cope with both cold start and node characteristics.In this dissertation,the recommendation algorithm based on graph neural network is studied(1)Aiming at the situation of anonymous users in real recommendation environment,this dissertation firstly studies the classification of anonymous nodes based on graph convolution network,and designs a unified feature processing mechanism in different anonymous environments.At the same time,aiming at the problem of bipartite heterogeneous graph reasoning in recommender system,this dissertation studies and designs a bipartite heterogeneous graph reasoning model based on Graph Auto Encoder,realizes the feature representation ability of bipartite graph nodes through cross domain message passing,and innovatively introduces the dual task learning mechanism,which effectively improves the reasoning ability of the model.(2)Aiming at the problem of "over smooth" feature in graph convolution neural networks,from the perspective of layer propagation,this dissertation introduces multilayer feature aggregation based on the previous work,which alleviates this problem to a certain extent and effectively improves the classification performance of the model.At the same time,from the point of view of reducing the complexity of the model,this dissertation introduces a cascade learning strategy,which optimizes the memory occupation of the model and exchanges the space optimization with less precision loss.(3)Based on the above works,this dissertation proposes a cold start dual task learning recommendation algorithm based on graph neural network.By analyzing the internal mechanism of the cold start problem,this dissertation determines the process flow with inductive learning as the core,and realizes the feature representation ability of the nodes outside the graph through the introduction of closed subgraph extraction and node re-labeling.On this basis,this dissertation proposes an innovative subgraph centric learning strategy,which predicts the "roles" of the nodes in the subgraph through the guidance model,in order to learn the topology of the subgraph to the greatest extent.On the other hand,by introducing dual task learning mechanism,this dissertation designs a collaborative optimization strategy based on central learning and score prediction to further improve the representation ability of the model.This dissertation validates the proposed model on three citation network datasets.The experimental results show better results compared to the baseline model and the existing mainstream methods.In addition,experiments are carried out on several mainstream recommendation system datasets.Compared with several state-of-the-art methods,the recommendation results of the proposed algorithm show significant advantages on these datasets.