[Paper Review] Learning under selective labels in the presence of expert consistency

Blum, Avrim, Stangl, Kevin|arXiv (Cornell University)|Jul 2, 2018

Fuzzy Systems and Optimization12 references17 citations

TL;DR

This paper proposes a data augmentation method to address selective labeling bias in algorithm-assisted decision making, where true outcomes are unobserved for cases where human experts consistently predict a certain label. By leveraging expert consistency—especially when humans are confident the outcome is negative—the method generates synthetic labels to improve model generalization. The key contribution is improved reliability and trustworthiness of models in selectively labeled data, particularly in detecting false negatives and reducing systemic discrimination risks.

ABSTRACT

Multiple fairness constraints have been proposed in the literature, motivated by a range of concerns about how demographic groups might be treated unfairly by machine learning classifiers. In this work we consider a different motivation; learning from biased training data. We posit several ways in which training data may be biased, including having a more noisy or negatively biased labeling process on members of a disadvantaged group, or a decreased prevalence of positive or negative examples from the disadvantaged group, or both. Given such biased training data, Empirical Risk Minimization (ERM) may produce a classifier that not only is biased but also has suboptimal accuracy on the true data distribution. We examine the ability of fairness-constrained ERM to correct this problem. In particular, we find that the Equal Opportunity fairness constraint [Hardt et al., 2016] combined with ERM will provably recover the Bayes optimal classifier under a range of bias models. We also consider other recovery methods including re-weighting the training data, Equalized Odds, and Demographic Parity, and Calibration. These theoretical results provide additional motivation for considering fairness interventions even if an actor cares primarily about accuracy.

Motivation & Objective

• To address the challenge of selective labeling bias in algorithm-assisted decision making, where true outcomes are unobserved for cases where human decisions block observation.
• To investigate whether expert consistency—especially in cases where humans are highly confident—can be used to inform model training when true labels are missing.
• To develop a data augmentation framework that incorporates expert certainty to improve model generalization to unlabeled regions of the data.
• To evaluate whether learning under selective labels leads to unreliable models prone to systemic discrimination, using the proposed method as an empirical validation tool.
• To explore the risks and benefits of relying on expert consistency when it may stem from shared biases rather than expertise.

Proposed method

• Proposes a data augmentation scheme that adds synthetic labels for instances where human decision probability P(di = 1|xi) < ϵ, assuming consistent human predictions indicate correctness.
• Uses a predictive model to estimate P(di = 1|xi), the probability that the true label will be observed, which is not subject to selective labeling bias.
• Applies a threshold ϵ = 0.05 to identify cases where humans are highly confident in not observing the outcome (i.e., di = 0), and assigns a semi-synthetic label based on expert consistency.
• Generates augmented training data SA = So ∪ {(xi, di) : P(di = 1|xi) < ϵ}, where So is the observed data, and uses this to train models on both observed and expert-consistent unlabeled cases.
• Employs inverse probability weighting in a variant to correct for residual selection bias in the augmented data.
• Validates the method using both real-world data (child welfare placement) and semi-synthetic data, comparing model performance on observed and augmented test sets.

Experimental results

Research questions

• RQ1 Can expert consistency in human decisions be reliably used to generate synthetic labels for unlabeled data under selective labeling?
• RQ2 Does data augmentation using expert consistency improve model generalization to the selectively unlabeled portion of the data?
• RQ3 To what extent does learning from selectively labeled data lead to unreliable models, especially in terms of false negative rates and systemic discrimination?
• RQ4 How does the model's performance on augmented data differ from performance on observed data, particularly in detecting high-risk cases?
• RQ5 Under what conditions is expert consistency indicative of correctness versus shared bias?

Key findings

• In semi-synthetic data, models trained only on observed data performed poorly at low sensitivity (high false negative rate), a flaw that was only detectable when evaluated on the augmented test set.
• The data augmentation method significantly reduced false negatives in semi-synthetic data, particularly for cases confidently screened out by experts, indicating improved generalization.
• In real-world data, the impact of augmentation was minimal due to low numbers of high-confidence screen-outs, but the model predictions showed strong agreement with human decisions in the screened-out region.
• The model trained on augmented data assigned lower risk scores to confidently screened-out cases, suggesting better alignment with expert consistency.
• The use of inverse probability weighting in the augmented setting did not significantly improve performance, indicating that the core benefit came from the expert-consistency-based augmentation itself.
• The method successfully incorporated expert knowledge into model training, enhancing trust and reliability in decision support systems where human confidence signals are informative.