Leveraging Explainability Maps for Group-unsupervised Robustness to Spurious Correlations

Abstract

Shortcut learning, the tendency for models to rely on spurious correlations, is a widespread issue in deep learning. Although being a known issue, uncovering the shortcuts present in a dataset can be a difficult task. Over the last few years, explainability methods have been leveraged to find previously unknown shortcuts within mainstream datasets. However, how to best mitigate a model’s reliance on shortcuts, is not a matter that is widely agreed on. Recently, the concept of group robustness has appeared as a potential way for mitigating shortcuts. In group robustness, the data in each class is divided into subclasses where some contain the shortcut features while some other subclass does not. By optimising a model to increase the worst group performance, the model learns to perform well across groups, mitigating reliance on shortcuts. One notable limitation that exists for current group robustness methods is their reliance on group labels to guarantee performance improvements. The issue with this is that acquiring these additional labels is a difficult and timeconsuming task. Therefore we propose, eXplainability-based Feature Reweighting (XFR), a group-unsupervised group robustness method. Our proposed method leverages the clustering of explainability heatmaps to estimate pseudolabels for groups in a dataset and afterwards uses these labels to improve group robustness. In our results, we show that XFR clearly improves group robustness compared to standardly trained models (ERM). We also show that performance is on par with, and sometimes even surpasses, methods that partially or fully utilise group labels.