Graph Regularization Attention Networks

Recently, graph attention network (GAT) has been widely used for classification and graph representation on graphstructured data. However, existing GAT fails to consider the consistency constraints in learning and representation process. In order to overcome this problem, we focus on using graph regularization attention network (GRAT) to learn better graph representation and more accurate classification of non-Euclidean data. The proposed GRAT model, on the one hand, encodes the graph structure and attention mechanism together, while naturally maintaining regularized attention; on the other hand, graph Laplace regularization ensures the local consistency constraint, so as to realize robust attention representation and classification. Comparative experiments on seven datasets show that GRAT has better classification performance.