Reliability Diagram

The core mechanism of a reliability diagram involves grouping predictions into bins based on their predicted confidence scores. Common strategies include:

• Equal-width bins: Dividing the confidence range [0,1] into fixed intervals (e.g., 10 bins of width 0.1).
• Equal-mass bins: Creating bins so each contains roughly the same number of prediction instances. The choice of binning strategy and the number of bins can affect the diagram's granularity and the interpretation of miscalibration patterns.

A diagonal reference line from (0,0) to (1,1) represents the ideal state of perfect calibration. For any given confidence level, the empirical accuracy should match the prediction. Points or bars lying above this line indicate underconfidence (the model is more accurate than it claims), while points below the line indicate overconfidence (the model is less accurate than its confidence suggests). This visual baseline is critical for rapid assessment.

For each bin, the diagram plots two key values:

• X-coordinate (Predicted Confidence): The average of the confidence scores for all predictions in that bin.
• Y-coordinate (Empirical Accuracy): The proportion of those predictions that were actually correct. A well-calibrated model will have points where the accuracy (y) equals the average confidence (x), causing them to fall along the perfect calibration line. Large deviations form a visible 'gap' representing miscalibration.

The diagram makes specific miscalibration patterns immediately apparent:

• Systematic Overconfidence: A curve that lies consistently below the diagonal, often seen in modern deep neural networks.
• Systematic Underconfidence: A curve that lies consistently above the diagonal.
• Non-Monotonic Miscalibration: A zig-zag pattern where the model is overconfident in some confidence ranges and underconfident in others, indicating more complex miscalibration that simple scaling cannot fix.

The Expected Calibration Error (ECE) is a scalar summary statistic directly derived from the reliability diagram. It is computed as the weighted average of the absolute vertical gaps between each bin's empirical accuracy and its average predicted confidence, with the weights being the proportion of samples in each bin. The reliability diagram provides the visual decomposition of the ECE, showing which confidence regions contribute most to the overall miscalibration score.

The primary use case for a reliability diagram is to validate the effectiveness of post-hoc calibration methods like temperature scaling or Platt scaling. Practitioners generate two diagrams: one for the uncalibrated model and one for the calibrated model. A successful calibration technique will shift the points or bars significantly closer to the diagonal line, providing visual proof that the confidence scores have been corrected. It is the standard diagnostic tool for comparing calibration techniques.

Expected Calibration Error (ECE) is the primary scalar metric derived from a reliability diagram. It quantifies miscalibration by computing a weighted average of the absolute difference between a model's average predicted confidence and its empirical accuracy across multiple bins.

• Calculation: ECE = Σ (|acc(b) - conf(b)| * n_b / N) across all bins.
• Interpretation: A lower ECE indicates better calibration. It provides a single number to track and compare models, but its value can be sensitive to the number of bins chosen.

Post-hoc calibration refers to techniques applied to a trained model's outputs—without retraining—to correct its confidence scores. A reliability diagram is used to diagnose the need for these methods and to validate their effectiveness.

• Common Methods: Include Temperature Scaling, Platt Scaling, and Isotonic Regression.
• Process: A held-out calibration set is used to fit simple scaling functions (e.g., a single temperature parameter) that map the model's original logits to better-calibrated probabilities.

Proper scoring rules are loss functions that measure the overall quality of probabilistic predictions, evaluating both calibration and sharpness. They provide a training signal for calibration-aware models and a final evaluation metric.

• Brier Score: Measures mean squared error between predicted probabilities and true binary outcomes. Lower is better.
• Negative Log-Likelihood (NLL): Penalizes models for assigning low probability to the correct class. It is the standard loss for training classification models and a proper scoring rule.

Calibration in production encompasses the MLOps practices required to maintain calibration after deployment. A reliability diagram generated on live data can reveal calibration drift due to changing data distributions.

• Monitoring: Requires periodic re-evaluation using a calibration pipeline that automatically scores new data, updates diagrams, and triggers retraining or recalibration.
• Challenge: Must address out-of-distribution (OOD) calibration, where models often become overconfident on unfamiliar inputs.

Multi-class calibration extends the binary concepts visualized in a reliability diagram to classification problems with more than two classes. The diagnostic becomes more complex, often focusing on the confidence of the top predicted class.

• Diagnosis: A reliability diagram can be constructed by binning the model's maximum predicted probability for each sample and comparing it to the accuracy of the top-class prediction.
• Methods: Techniques like matrix scaling (a multi-class extension of Platt scaling) are used for post-hoc correction.

Selective calibration is a strategy where a model is allowed to abstain from making predictions on low-confidence inputs. The goal is to maintain high calibration only on the subset of instances for which it does predict, increasing trust in its active outputs.

• Relationship to Reliability: A reliability diagram for a selective model would only include bins for confidence levels above the abstention threshold.
• Use Case: Critical in high-stakes applications like medical diagnosis or autonomous systems, where wrong but confident predictions are unacceptable.