Research On Personalized Recommendation Of Social Networks Based On Graph Convolution

In recent years,with the shift of Internet information from the era of dramatic increase in data volume to the era of stock,in the field of recommendation search,in addition to accurately predicting the items that users are most likely to like,tapping into users’ potential interests is also an aspect that needs to be considered.By considering diversity and accuracy to solve the problem of repeated recommendation of items in the same category,it can not only effectively reduce the user’s fatigue of information reading,but also effectively alleviate the information cocoon problem in the recommendation system.Regarding the problem of information redundancy and longtail distribution brought by huge amount of data,it needs to be solved by considering diversification when generating recommendation item collections.The essence of the recommendation system is to serve users,and in order to meet the personalized needs of different users,two aspects of user satisfaction need to be considered.On the one hand,the results of the recommendation system should be improved in terms of accuracy,and on the other hand,the diversity of recommendation results should be improved.In this paper,the graph neural network of social relationship is constructed by using graph convolution structure,and the information of each node is passed and aggregated by multi-layer neural network,which makes full use of the high-order potential information in the neural network and solves the problem that a small amount of data cannot effectively predicted.On the basis of considering the accuracy,in order to enhance the effect of diverse recommendations,this paper explores the following:(1)Firstly,a personalized recommendation algorithm model based on graph neural network is proposed,which is based on the graph convolution framework,building a graph neural social network by stacking multiple layers of structural information,allowing information transfer and aggregation between different levels of neural networks,each node get effective embedding information and training parameters.(2)In this paper,we use multilayer graph neural network to regulate the node embedding vectors,which will greatly reduce the recommendation efficiency if we use the last layer embedding vectors after propagation as the embedding information of nodes.By introducing the layer attention adjustment mechanism to obtain the embedding vector information of nodes in different layers,the hyperparameters selfadjust the weights of different layers instead of the fixed weights in the light graph convolution during the training process,which can effectively alleviate the oversmoothing problem.(3)Due to the large amount of data,neighbor sampling is selected to construct subgraphs for training.To address the problem of small coverage of recommendation results due to the long-tail distribution of data and the concentration of recommendation results on a few kinds of items,we propose the method of diverse neighbor sampling to obtain as many kinds of item choices as possible by maximizing the diversity of neighbors in the subgraph when constructing the subgraph.(4)For the selection of the loss function,we use the BPR loss function to count the losses,and a small number of categories contain a large number of items due to the long-tail distribution of items.this paper proposes category-based loss weights for different categories of items plus the category weight parameter to calculate the loss.(5)In order to make the recommendation results more accurate and faster training,this paper chooses the negative sampling method,and proposes two methods to improve negative sampling compared with random negative sampling,one is to negatively sample in the set of user-independent items,and the other is to negatively sample in the set of items in the same category of interactive items but without interaction.(6)The proposed model is compared with other classical recommendation algorithms on the same data and the final results of the experiments confirm the validity of the proposed model.