Graph Representation Learning

Graph Representation Learning uses graph data structures as inputs to Machine Learning algorithms. While it is possible to naively process every node in a graph as a single data point, standard techniques should not necessarily be applied to graphs, since this ignores a key feature of graph representation: nodes are not independent data points, they are interconnected. The edges between nodes are a fundamental part of the graph, and must be processed by the algorithm.

Both supervised and unsupervised learning techniques may be used with graphs (data for nodes and edges, or just shape). There are four types of prediction tasks that can be performed on a graph:

• →

  Predict the features or properties of a node (e.g. classify it)

• →

  Predict the features or properties of an edge (link between two nodes)

• →

  Detect communities or clusters of nodes within a graph

• →

  Make a prediction about or classify the entire graph