foreword

Our world is highly rich in structure, comprising objects, their relations, and hierarchies. Sentences can be represented as sequences of words, maps can be broken down into streets and intersections, the world wide web connects websites via hyperlinks, and chemical compounds can be described by a set of atoms and their interactions. Despite the prevalence of graph structures in our world, both traditional and even modern machine learning methods struggle to properly handle such rich structural information: machine learning conventionally expects fixed-sized vectors as inputs and is thus only applicable to simpler structures such as sequences or grids. Consequently, graph machine learning has long relied on labor-intensive and error-prone handcrafted feature engineering techniques. Graph neural networks (GNNs) finally revolutionize this paradigm by breaking up with the regularity restriction of conventional deep learning techniques. They unlock the ability to learn representations from raw graph data with exceptional performance and allow us to view deep learning as a much broader technique that can seamlessly generalize to complex and rich topological structures.