Deep Recommendation System Based On Social Network Graph Convolution

In recent years,with the continuous development of big data technology,represented by online shopping platforms,the online platform pays more and more attention to the use of recommendation systems.The most critical component of the recommendation system is the personalized recommendation algorithm.The traditional recommendation algorithm needs to rely on a large number of historical data for model training,so as to improve the recommendation effect.While solving the problem of information overload in current network data,the recommendation system still has the following problems: first,the problem of data sparsity.Even if the historical information of all users is very rich,there will be very few historical records of some users.In the large-dimensional interaction matrix,the problem of data sparsity is prominent.The second,the cold start problem.The recommendation system will continue to add new users and items for the recommendation.How to improve the recommended content of newly added items has become a new problem.Third,scalability,whether the recommendation system can maintain the previous recommendation efficiency in the face of the surge of data.In order to meet these challenges,this dissertation does the following research:(1)In order to avoid the reduction of recommendation efficiency caused by data sparsity and cold start,combined with the user’s social relationship,this dissertation extracts the information conducive to the recommendation,and generates the user’s representation vector according to the social graph structure.(2)This dissertation proposes a recommendation model based on user social graph convolution,which uses a graph convolutional neural network to aggregate the information of neighbor nodes in the user’s social network.The two-layer graph convolutional neural network can aggregate users according to their user sequence.The information of second-order neighbors generates the user’s social representation vector.(3)The user appears in both the user’s historical interaction graph and the social graph in the social recommendation.This dissertation combines the user’s special identity in the social recommendation and divides social recommendation into user-item interaction,social graph,and item-There are three parts of user interaction.The user-item interaction part uses long short-term memory neural networks to aggregate the information of the user’s historical interaction sequence.The social graph uses a graph convolution neural network is used to aggregate the user’s neighbor information.The user-item interaction uses the user’s social representation aggregated by the social graph as the query vector of the attention mechanism.The attention mechanism allocates the weight according to the characteristics of the user’s neighbor nodes.(4)This dissertation conducts a lot of comparative experiments on two real data sets for the proposed model.First,the data of the dataset is preprocessed,and secondly,by comparing with multiple baseline models,the model proposed in this paper can effectively improve the performance of the recommendation system.Finally,the three parts of the model in this dissertation are the best through ablation experiments.