MC-GAT: Multi-Channel Graph Attention Networks for Capturing Diverse Information in Complex Graphs

Graph attention networks (GAT), which have strong performance in tackling various analytical tasks on network data, have attracted wide attention. However, complex real-world networks have both edge topology and node features. GAT only relies on the topology of edges to extract network information, and the association between node features is underutilized, which may seriously hinder GAT’s expressive ability on some tasks. In addition, the attention mechanism can automatically assign different weights to different pieces of information, making it easier to express information with multiple aspects. Therefore, we propose semi-supervised multi-channel attention networks (MC-GAT), which simultaneously extract node features, topological structures, and their combination information. The MC-GAT model consists of two specific attention modules, one common attention module, and the attention mechanism. To create node embeddings containing various informational aspects, we use the attention mechanism to assign weights to each. Extensive testing on benchmark datasets has shown us to be at our best. The performance of the proposed model is demonstrated by the fact that MC-GAT achieves relative maximum improvements of 4.22% for accuracy (ACC) on BlogCatalog and 5.23% for macro F1-score (F1) on UAI2010. Experimental results on relevant datasets show that the method has satisfactory performance, and multi-channel graph attention can capture richer structural and feature information within linear time complexity. This work provides a new way of thinking about graph neural networks.