Research On Personalized Recommendation System Based On Graph Embedding Representation

Recommendation system is an important means to solve the problem of information overload,which can filter useless information and improve the efficiency of information retrieval.By mining the user’s historical behavior and personal attributes,it models the user’s interests,and then finds the most suitable items to recommend to the user in the massive project collection.Graph embedding representation learning technology has been widely used in the algorithm model of recommendation system at present.This paper focuses on three problems of matrix factorization and graph neural network in graph embedding representation learning,and proposes three corresponding personalized recommendation models.The specific content is as follows:First of all,aiming at the problem that the dot product attribute in the matrix factorization recommendation algorithm limits the expression of similarity between items and users,leading to sub-optimal solutions,a recommendation model based on interactive neighborhood distance factorization(NDF)is proposed.Based on the distance factorization algorithm,the model uses user-item interaction neighborhood information to reconstruct the spatial location features of users and items,and introduces an attention mechanism into the reconstructed user location features to obtain more accurate user preferences.The experimental results show that NDF has better accuracy and robustness in rating prediction tasks than other models.Secondly,aiming at the problem that the recommendation model based on graph convolutional neural network cannot dynamically assign weights to neighbor nodes,a matrix completion recommendation model based on graph attention network(MCGA)is proposed.The model consists of a graph attention encoder and a bilinear decoder.The graph attention encoder automatically learns the importance of different neighbor nodes by introducing a multihead attention mechanism into the graph encoder.The bilinear decoder implements the matrix completion task in the form of link prediction.In node coding,the model constrains the spatial complexity of the model by learning the potential features of low-dimensional users and projects as inputs.The experimental results validate the effectiveness of the multi-head attention mechanism in MCGA.Finally,aiming at the bottleneck problem caused by the information aggregation of input nodes containing only ID information in the graph convolutional network,the introduction of text reviews as auxiliary information is considered,and a lightweight graph convolutional network recommendation model incorporating text information(T-LGCN)is proposed.The model uses convolutional neural network to extract text comment features and merge them into graph network nodes.Use the simplified lightweight graph convolutional network to linearly propagate the fusion features of neighbor nodes on the user-item interaction graph to learn user and item embeddings,and use the weighted sum of embeddings learned on all layers as the final embedding.The rating prediction is completed in the form of dot product.The experimental results validate that T-LGCN can effectively alleviate the bottleneck of graph convolution information aggregation,and at the same time alleviate the user’s cold start problem to a certain extent.