A Component-level Attention based Adaptive Graph Convolutional Network

Graph neural networks have attracted more and more attention in the task of learning node representations. Currently, most graph neural networks are usually applied to assortative graphs. They can't perform well in disassortative graphs, where adjacent nodes tend to have different class labels. However, previous studies have shown that it is effective for nodes to gather different information from their neighbors not only in disassortative graphs but also in assortative graphs. Based on this insight, this paper proposes a novel Component-level Attention Adaptive Graph Convolutional Network (CAAGCN) to collect different information more efficiently. Firstly, the model collects dissimilar features among nodes by introducing a component-level attention mechanism. One attention component learns the importance of neighbor nodes and the other regulates the proportion of the similar characteristics of neighbor nodes. Secondly, in order to collect different features between nodes more effectively, we optimize the input function during the attention learning process and preprocess the node features. It can effectively alleviate over-smoothing. Finally, extensive experiments on six networks verify the effectiveness of the method in node classification task.