A Survey on GAT-like Graph Neural Networks

Abstract

The graph structure is one of the critical data structures in the real world, and its applications focus on graphs, where scholars study entity features and interactions among various entities. Recently, developments in graph neural networks (GNNs) have heightened the need for learning graph representations effectively. Simultaneously, graphs can be large and complex as well as noisy, posing obstacles for graph-related tasks. However, by incorporating the attention mechanism in graph neural networks, it is possible for GNNs to focus on the most important entities and interactions in graphs, contributing to better decisions. Therefore, this paper conducts a comprehensive survey about literature on GAT-like graph neural networks. According to inputs and outputs, types of attention mechanisms, tasks, this paper proposes a taxonomy to group recent works followed by detailed examples, aiming to overlook GAT-like GNNs from different perspectives. At last, this paper discusses the existing problems and challenges in this area, hoping to provide insights for future research directions.