Learning Concept Credible Models for Mitigating Shortcuts

Published in NeurIPS 2022, 2022

Recommended citation: Jiaxuan Wang, Sarah Jabbour, Maggie Makar, Michael W. Sjoding, Jenna Wiens. Learning Concept Credible Models for Mitigating Shortcuts. *NeurIPS*, November 2022. https://openreview.net/pdf?id=yKYCwTvl8eU

During training, models can exploit spurious correlations as shortcuts, resulting in poor generalization performance when shortcuts do not persist. In this work, assuming access to a representation based on domain knowledge (i.e., known concepts) that is invariant to shortcuts, we aim to learn robust and accurate models from biased training data. In contrast to previous work, we do not rely solely on known concepts, but allow the model to also learn unknown concepts. We propose two approaches for mitigating shortcuts that incorporate domain knowledge, while accounting for potentially important yet unknown concepts. The first approach is two-staged. After fitting a model using known concepts, it accounts for the residual using unknown concepts. While flexible, we show that this approach is vulnerable when shortcuts are correlated with the unknown concepts. This limitation is addressed by our second approach that extends a recently proposed regularization penalty. Applied to two real-world datasets, we demonstrate that both approaches can successfully mitigate shortcut learning.