From Biased Selective Labels to Pseudo-Labels: An Expectation-Maximization Framework for Learning from Biased Decisions
This addresses the issue of biased decision-making in machine learning for clinical applications, offering a method to mitigate labeling biases that could otherwise amplify disparities, though it is incremental as it builds on existing causal models.
The paper tackles the problem of learning from biased selective labels, where labeling biases vary across subgroups and unlabeled cases are imputed as negative, by proposing the Disparate Censorship Expectation-Maximization (DCEM) algorithm. It shows that DCEM improves bias mitigation (area between ROC curves) without sacrificing discriminative performance (AUC) on synthetic and clinical sepsis classification data.
Selective labels occur when label observations are subject to a decision-making process; e.g., diagnoses that depend on the administration of laboratory tests. We study a clinically-inspired selective label problem called disparate censorship, where labeling biases vary across subgroups and unlabeled individuals are imputed as "negative" (i.e., no diagnostic test = no illness). Machine learning models naively trained on such labels could amplify labeling bias. Inspired by causal models of selective labels, we propose Disparate Censorship Expectation-Maximization (DCEM), an algorithm for learning in the presence of disparate censorship. We theoretically analyze how DCEM mitigates the effects of disparate censorship on model performance. We validate DCEM on synthetic data, showing that it improves bias mitigation (area between ROC curves) without sacrificing discriminative performance (AUC) compared to baselines. We achieve similar results in a sepsis classification task using clinical data.