Graph Neural Network model calibration for trusted predictions

Graph neural networks (GNNs) are a fast developing machine learning specialisation for classification and regression on graph-structured data. They are a class of powerful representation learning algorithms that map the discrete structure of a graph, e.g., nodes and edges, to a continuous vector representation trainable via stochastic gradient descent.

These representations can be used as input to classification and regression algorithms targeting a variety of applications including finance, genomics, communications, transportation and security. But when applying new machine learning models to real-world problems, we must ask the question: how reliable are they?

In this article, we’ll talk about calibration in graph machine learning, and how it can help to build trust in these powerful new models.

For a detailed overview of graph machine learning and its applications read Knowing your Neighbours: Machine Learning on Graphs

Classification for graph data

This discussion will focus on only the classification setting for graph data. We consider the problem of predicting a discrete label (binary or multi-class) for the nodes of a graph, given we have observed the labels for a subset of the nodes, their attributes and the…