Decoupled Graph Attention Networks

Abstract

Graph Neural Networks (GNNs) have garnered significant attention due to their ability to learn graph representations. Recent years have seen research interests shift toward the exploration of parameter-efficient algorithms for graph machine learning. Existing frameworks designed under the assumptions of homophily showcase limited performance on heterophilic graphs. In this paper, the challenge of heterophily is tackled while reducing excess complexity observed in standard message-passing frameworks by reducing excessive non-linear transformations and introducing pre-computable primitives. The evaluation of Decoupled Graph Attention Networks (DGAT) on various heterophilic graphs portrays a significant improvement of up to 1.31 times compared to standard baseline approaches.