Graph-based AI

Graphs open up opportunities to represent complex relationships between entities for describing abstractly semantics of a real-world use case. There are numerous application areas: Molecules in biology, traffic systems, virtual communities, business ecosystems, etc. etc. Network analysis metrics, e. g., betweenness centralities of nodes or clustering coefficients of graphs, can serve as input features for machine learning (ML) models, if there is a lack of data. The current research stream of graph neural networks (GNN) can solve several problems like link prediction, node and graph classification. For example, the behavior of different molecules can be seen as graph classification task, the prediction of friendships on social networks applies link classification, and warning systems for traffic density are based on node classification. Graphs differ from traditional machine learning data input such as numbers, text or images. As graphs come with different sizes and unordered nodes, they are more difficult to process. The basic idea is to find algorithms so that each node can be represented by its direct neighbors and their attributes. These embeddings are then fed into a classical neural network. As many real-world problems can be expressed as graphs, this field shows a high potential for applied research.