Towards OOD Detection in Graph Classification from Uncertainty Estimation Perspective

TL;DR

This paper investigates out-of-distribution detection in graph classification from an uncertainty estimation perspective, comparing recent methods and highlighting the importance of graph representations and predictive distributions.

Contribution

It provides a comparative analysis of OOD detection methods for graph classification, emphasizing the need to consider both graph representations and predictive distributions.

Findings

No universal OOD detection approach for graphs

Importance of graph representations in OOD detection

Significance of predictive categorical distribution

Abstract

The problem of out-of-distribution detection for graph classification is far from being solved. The existing models tend to be overconfident about OOD examples or completely ignore the detection task. In this work, we consider this problem from the uncertainty estimation perspective and perform the comparison of several recently proposed methods. In our experiment, we find that there is no universal approach for OOD detection, and it is important to consider both graph representations and predictive categorical distribution.