Selective Deferral Under Subgroup Shift

Research Question

Design a practical selective prediction and deferral policy for high-stakes tabular decisions.

The task isolates one modular question: given a fixed base classifier, what acceptance / deferral rule best trades off selective risk, subgroup fairness, and overall discrimination?

Background

Selective prediction systems should be able to say "I do not know" when the classifier is uncertain. In high-stakes settings, that deferral can be handed to a human reviewer or a slower backup process. The benchmark studies whether a policy can:

• keep selective risk low at a fixed target coverage,
• avoid concentrating deferrals on one subgroup,
• preserve AUROC as a confidence ranking signal, and
• remain simple enough to run offline on modest compute.

Task

Modify the SelectivePolicy class in custom_selective.py. The rest of the pipeline is fixed: dataset loading, train / calibration / test splitting, base model training, and metric computation.

The policy receives calibration-time base-model probabilities and subgroup labels, then decides whether each test example should be accepted or deferred. You may implement a single global threshold, a learned deferral score, subgroup-specific thresholds, or any other compact policy that fits the interface.

Evaluation

The benchmark runs on four offline tabular proxies from scikit-learn:

• breast_cancer
• diabetes (binarized around the training-set median)
• california (binarized around the training-set median)
• madelon

Each dataset is split into train / calibration / test partitions. Subgroups are formed from a stable feature threshold so that worst-group behavior can be measured.

Metrics:

• selective_risk_at80: classification error on accepted examples at 80% target coverage
• worst_group_selective_risk: worst subgroup error on accepted examples
• deferral_rate_gap: max subgroup deferral rate minus min subgroup deferral rate
• auroc: AUROC of the acceptance score for predicting correctness

Baselines

• confidence_thresholding: tune one confidence threshold to hit the target coverage
• conformal_abstention: split-conformal abstention with a coverage target
• learned_deferral: train a compact meta-model that predicts whether the base model will be correct
• groupwise_thresholding: subgroup-specific thresholds as a stronger reference baseline

Practical Notes

This task intentionally uses datasets that are already available offline in the repository's scikit-learn package setup, so no shared package edit is required.

If you want the exact adult / ACSIncome / COMPAS / Law School datasets from the prompt, that would require a shared package-data follow-up, which I have not implemented here.