Selective Classification

The core mechanism that allows a model to abstain from making a prediction when its confidence falls below a predefined threshold. Unlike standard classifiers forced to guess, a selective classifier outputs 'I don't know' for ambiguous or unfamiliar inputs.

• Hard Abstention: Binary decision—predict or reject entirely
• Soft Abstention: Defer to a human expert or fallback system
• Threshold Tuning: The rejection threshold is calibrated on a held-out validation set to balance coverage (percentage of inputs predicted on) against risk (error rate on accepted predictions)

The mathematical function that quantifies a model's certainty in its prediction. The quality of selective classification depends entirely on how well this score separates correct from incorrect predictions.

• Maximum Softmax Probability (MSP): The simplest baseline—uses the highest class probability as confidence. Prone to overconfidence on OOD inputs
• Entropy of Predictive Distribution: Measures uncertainty across all classes; high entropy signals confusion
• Energy-Based Score: Derived from the logit values using the Helmholtz free energy formula. Superior at out-of-distribution detection compared to MSP
• Monte Carlo Dropout Variance: Uses variance across multiple stochastic forward passes as an epistemic uncertainty proxy

The fundamental operating curve of any selective classifier, plotting error rate against coverage. As the model rejects more inputs (lower coverage), the risk on accepted predictions should monotonically decrease.

• Area Under the Risk-Coverage Curve (AURC): A scalar summary metric; lower AURC indicates better selective performance
• Coverage at Fixed Risk: A practical metric—what percentage of inputs can the model handle while guaranteeing a target error rate (e.g., 95% coverage at 1% risk)
• Selective Accuracy Constraint: The model must achieve a specified accuracy on accepted predictions, even if it means rejecting a large fraction of inputs

Selective classification performance is directly dependent on the calibration quality of the underlying confidence scores. A miscalibrated model will make poor abstention decisions.

• Overconfident Models: Produce high confidence scores even when wrong, causing the reject option to fail at filtering errors
• Underconfident Models: Reject too many correct predictions, unnecessarily reducing coverage
• Post-hoc Calibration: Techniques like temperature scaling or Platt scaling are often applied before setting the rejection threshold to align confidence with empirical accuracy
• Proper Scoring Rules: Training with Brier score or negative log-likelihood encourages inherently better-calibrated confidence estimates

Two related but distinct frameworks for prediction with abstention. Understanding their differences is critical for selecting the right approach.

• Selective Classification: Relies on a confidence score and a fixed threshold. Provides no formal guarantee on error rate; performance is empirical and depends on calibration quality
• Conformal Prediction: A distribution-free framework that produces prediction sets with a finite-sample, marginal coverage guarantee. Instead of a single class, it outputs a set containing the true label with user-specified probability (e.g., 90%)
• Hybrid Approaches: Conformal risk control can be applied to selective classification to provide statistical guarantees on the false negative rate of accepted predictions

How selective classification is operationalized in production systems, particularly in high-stakes domains where the cost of an error far exceeds the cost of abstention.

• Human-in-the-Loop Triage: The model predicts on high-confidence cases and routes rejected instances to human reviewers. Common in medical imaging and legal document review
• Fallback Cascade: Rejected predictions cascade to a more expensive but more accurate model (e.g., a large ensemble or a human expert)
• Safety-Critical Guardrails: In autonomous systems, abstention triggers a safe default action rather than a risky autonomous decision
• Cost-Sensitive Thresholding: The rejection threshold is set by minimizing expected cost, where false positives, false negatives, and abstention each carry distinct economic weights

The primary metric for diagnosing if a model's confidence scores are trustworthy. ECE partitions predictions into M equal-width bins and computes the weighted average of the absolute difference between accuracy and confidence within each bin. A low ECE is a prerequisite for setting a meaningful abstention threshold; without it, a selective classifier cannot reliably map a confidence score of 0.9 to a true 90% likelihood of correctness.

The central diagnostic tool for evaluating a selective classifier. This plot visualizes the trade-off between the error rate (risk) on accepted predictions and the proportion of inputs predicted on (coverage). As the confidence threshold increases, coverage drops but accuracy rises. The Area Under the Risk-Coverage Curve (AURC) provides a single scalar summary metric; a lower AURC indicates a superior ability to identify and abstain from likely errors.

A post-hoc calibration method that acts as a precision dial for a model's softmax outputs. A single scalar parameter T divides all logits before the softmax function:

• T > 1: Softens probabilities, reducing overconfidence.
• T < 1: Sharpens probabilities, increasing peakiness. Optimized on a held-out validation set to minimize Negative Log-Likelihood (NLL), this parameter is critical for selective classification because it directly controls the raw confidence scores fed into the abstention decision gate.

A distribution-free framework that wraps any pre-trained model to produce prediction sets with a finite-sample, marginal coverage guarantee. Instead of a single label, it outputs a set containing the true label with a user-specified probability (e.g., 95%). For selective classification, the size of this set acts as an uncertainty metric: a singleton set signals high confidence, while a large or empty set triggers abstention. Unlike Platt scaling, it requires no distributional assumptions.

The task of identifying inputs that are semantically alien to the training distribution. Selective classification and OOD detection are complementary rejection strategies:

• Selective Classification: Abstains on in-distribution inputs where the model is uncertain.
• OOD Detection: Abstains on inputs from a fundamentally different domain. Modern systems combine both, using energy-based models or Mahalanobis distance scores to first filter OOD inputs before applying a calibrated confidence threshold to the remaining in-distribution data.

Loss functions that are mathematically minimized only when the predicted probability distribution exactly matches the true data-generating distribution. They penalize dishonesty and reward perfect calibration. Key examples include:

• Brier Score: Mean squared error between predicted probability and outcome.
• Negative Log-Likelihood (NLL): Heavily penalizes confident misclassifications. Training with a proper scoring rule encourages the calibrated probability estimates that a selective classifier's abstention logic depends on.