A survey of graph neural networks in various learning paradigms: methods, applications, and challenges

In the last decade, deep learning has reinvigorated the machine learning field. It has solved many problems in computer vision, speech recognition, natural language processing, and other domains with state-of-the-art performances. In these domains, the data is generally represented in the Euclidean space. Various other domains conform to non-Euclidean space, for which a graph is an ideal representation. Graphs are suitable for representing the dependencies and inter-relationships between various entities. Traditionally, handcrafted features for graphs are incapable of providing the necessary inference for various tasks from this complex data representation. Recently, there has been an emergence of employing various advances in deep learning for graph-based tasks (called Graph Neural Networks (GNNs)). This article introduces preliminary knowledge regarding GNNs and comprehensively surveys GNNs in different learning paradigms-supervised, unsupervised, semi-supervised, self-supervised, and few-shot or meta-learning. The taxonomy of each graph-based learning setting is provided with logical divisions of methods falling in the given learning setting. The approaches for each learning task are analyzed from theoretical and empirical standpoints. Further, we provide general architecture design guidelines for building GNN models. Various applications and benchmark datasets are also provided, along with open challenges still plaguing the general applicability of GNNs.

Automated summary

This article introduces recent advances in applying deep learning to graph-based tasks, known as Graph Neural Networks (GNNs). It covers different learning paradigms, including supervised, unsupervised, semi-supervised, self-supervised, and few-shot or meta-learning. The methods for each learning task are analyzed from theoretical and empirical perspectives, and general guidelines for building GNN models are provided, along with applications and benchmark datasets.