Research On Recommendation System Model Based On Graph Network Structure

As the impact of "information overload" on the Internet expands,recommendation systems have become an integral part of people’s lives.The quest for higher performance recommendation systems and model algorithms by personalized service providers has also attracted a lot of attention from researchers in academia and industry.Graph neural networks(GNNs)have advantages in data processing on graph domains,which can naturally integrate node neighborhood information and topology,and have been successfully applied to recommendation problems in recent years.At present,GNN-based recommendation models mainly have the following shortcomings: the first point is the direct use of graph convolutional neural networks(GCNs)in user-item bipartite graphs without differentiating the neighbor nodes in the neighborhood,which are different in the category of nodes in the neighborhood on the one hand,and the connection tightness between the central node and the neighbor nodes on the other hand.The second point is that most graph convolutional network models do not pay attention to the internal order of historical interaction sequences for user interest feature extraction,and thus cannot distinguish between long-term and short-term interests of users.The third point is that it is difficult for a single GCN network to learn the embedding representation of nodes in a graph under different feature views with high quality.To address the above problems,the main contributions of this paper to the recommendation model based on graph network structure are as follows.(1)The construction from user-rating bipartite graph to user-homogeneous graph and item-homogeneous graph is completed by using SVD matrix decomposition.Then,the classical GCN model is improved and a weighted graph convolutional neural network is proposed and applied to collaborative filtering recommendation considering entity similarity,where potential features of users and items are extracted on two weighted networks respectively,and the binary vectorized entity features are embedded using one-hot as well as muti-hot coding.Finally,the recommendation task of rating prediction is completed using vector inner product,and the performance tests on RSME and MAE evaluation metrics are completed on both datasets.The results show that our model shows some superiority over the traditional matrix decomposition class model and the classical GCN class model.(2)A collaborative filtering recommendation model(LGACF)combining LSTM and graph attention network is proposed by dividing users’ interests into short-term and long-term preferences.For user short-term preference features,the model invokes the long and short-term memory network and the self-attention mechanism to make it adaptively assign weights to the user’s historical interaction items by giving different degrees of attention to the self-attention part during the training process,and the long and short-term memory network is used to learn the internal dependencies of the user’s historical interaction sequences to obtain the user short-term interest representation.(3)For user long-term preferences,LGACF also includes a multi-view graph attention network proposed in this paper,which uses multiple graph attention networks to adaptively learn and aggregate the node features under the rating view and feature view on the bipartite graph of user items,and obtains the long-term interest representation and item representation of users.Finally,the Top-K recommendation task was completed using vector inner product,and the performance was tested on HR@K and NDCG@K evaluation metrics in both datasets.The results show that LGACF has a better performance compared to the sequence-based recommendation model and the graph convolution class recommendation model.