Graph convolution neural network for recommendation using graph negative sampling
After several years of rapid development,the collaborative filtering algorithms based on graph convolutional neural networks have achieved the most advanced performance in many recommender system scenarios.However,most of these algorithms only use simple random negative sampling method when collecting negative samples,and do not make full use of graph structure information.To solve this problem,a graph convolution neural network for recommendation using graph negative sampling(GCN-GNS)is proposed.The algorithm first constructs a user-item bipartite graph and uses a graph convolution neural network to obtain the node embedding vector.Next,the depth-first random walk method is used to obtain the sequence of the wandering item nodes that includes both the neighboring item nodes and the distant item nodes.Then the attention layer is designed to learn the weights of different nodes in the walk sequence adaptively and a dynamically updated virtual negative sample is formed according to the weights.Finally,the virtual negative sample is used to train the model more efficiently.Experimental results show that the GCN-GNS performs better than other algorithms for comparison on three real public datasets in most cases,which indicates that the proposed novel graph negative sampling method can help the GCN-GNS model to make better use of the graph structure information,and ultimately improves the effect of item recommendation.
万方数据
维普
