Chapter 4 - Graph convolutional networks

Abstract

Deep learning approaches have been very successful in many machine learning tasks including compute vision, natural language processing, audio processing, and speech recognition. However, deep neural networks typically work with grid-structured data represented in the Euclidean space and despite their recent successes, they poorly generalize to applications where the data is represented in non-Euclidean space. Recently, due to the increasing amount of graph structured data produced in different areas such as social networks, stock markets, and knowledge bases, there is an increasing need to develop learning methods capable of capturing the relational structures in graph data. Recently, graph neural networks have attracted great research attention as they have demonstrated the ability to provide high performance in many of the learning tasks with non-Euclidean data structures. In this chapter, we provide a detailed overview of graph convolutional networks, which extend the convolution operation to graph structured data. We group the existing graph convolutional networks in four different categories, and we provide a discussion on their efficiency and scalability in large graph structures. The application of these networks in different learning tasks is also discussed in this chapter, along with a summary of the existing benchmark data sets and open-source libraries.