A Systematic Survey of Graph Convolutional Networks for Artificial Intelligence Applications

Abstract

Graph Convolutional Networks (GCNs) have become an essential tool for handling graph‐structured data, enhancing the functionality of conventional convolutional neural networks (CNNs) in non‐Euclidean contexts. GCNs are particularly proficient in tasks such as node classification, link prediction, and graph clustering by collecting information from neighboring nodes. These models are utilized in a range of domains, including recommendation systems, social network analysis, bioinformatics, and computer vision. GCNs demonstrate significant effectiveness in challenges like citation prediction and knowledge graph completion, where both the structure of the graph and the information from the nodes are crucial. Emerging from the field of graph signal processing, GCNs have been enhanced by a variety of models that combine spectral and spatial convolution methods. Despite these improvements, there remain obstacles to fully harnessing the structural information of graphs, which is a vital component of network science. This survey presents an extensive review of GCNs and introduces a new taxonomy that classifies models into five categories: supervised, unsupervised, semi‐supervised, weakly‐supervised, and self‐supervised GCNs. We emphasize recent innovations, discuss present challenges, and propose promising avenues for future investigations.