Spurious associations in training data arise because models learn correlations rather than causal relations. We propose a composite loss that incorporates causal and behavioral priors: an invariance penalty for label-preserving perturbations (e.g., swapping gendered pronouns should not change predictions), a directional penalty for perturbations that should change the label in a known direction, and a falsification penalty inspired by epidemiological negative controls that discourages reliance on features with no causal relationship to the outcome. The framework unifies ideas from CheckList-style behavioral testing, causal inference, and adversarial robustness into a single training objective.
Cite: Gaurav Sood. "Causal Debiasing for Robust Machine Learning." https://gsood.com/research/causal-debiasing-for-robust-machine-learning/
BibTeX:
@misc{soodcausal,
author = {Gaurav Sood},
title = {{Causal Debiasing for Robust Machine Learning}},
note = {Working paper},
url = {https://gsood.com/research/causal-debiasing-for-robust-machine-learning/}
}