Mahalanobis Distance-Based Graph Attention Networks

Abstract

In this paper, a Mahalanobis Distance-based Graph Attention Network for graph classification, is proposed. In contrast to traditional Graph Attention Networks, the proposed approach learns the covariances between node features so as to determine the attention between nodes, instead of directly learning the attention coefficients using learnable parameter matrices. During training, the network learns the covariance matrix that is essential component of the Mahalanobis distance, and thus adjusts the covariances between node features based on the specific characteristics of the graph under examination. Leveraging Mahalanobis distance, the model manages to capture complex features correlations leading to better graph representations. Additionally, the proposed method combines the concepts of multi-head and multi-view to achieve enhanced performance and generalization ability. Multi-head attention enables the model to focus on diverse aspects of the data, whereas multi-view attention provides different perspectives on node relationships. Extensive experiments on benchmark datasets demonstrate that the proposed method either outperforms or is on par with the state-of-the-art methods. The study also examines the impact of the number of heads and views for the multi-head and multi-view concepts respectively on the proposed method.