Evaluation

Offline Evaluation

Golden Set Curation

  • Criteria for selection

    • Coverage: Includes all relevant feature distributions.

    • Accuracy: Labels verified by experts.

    • Diversity: Edge cases, rare conditions.

  • Update frequency?

    • Periodically (e.g., quarterly) or when drift is detected.

  • How to balance representation?

    • Maintain real-world distribution while oversampling rare cases.

Metric

  • Basic Building Blocks

    • TP = True Positives

    • FP = False Positives

    • FN = False Negatives

    • TN = True Negatives

  • From these:

    • Precision = TP / (TP + FP)

    • Recall = TP / (TP + FN)

    • F1-score = 2 × (Precision × Recall) / (Precision + Recall)

Binary Classification

  • ROC-AUC: Measures ability to distinguish classes across all thresholds; useful when class balance is not extreme.

  • PR-AUC: Focuses on positive class performance (precision vs recall); useful when positives are rare.

  • When to prefer ROC-AUC vs PR-AUC?

    • ROC-AUC: When positives and negatives are balanced.

    • PR-AUC: When positives are rare (e.g., fraud detection, rare disease prediction).

Multi-class Classification (1 label/sample)

  • \(n_i\) is the number of true instances for class i, and \(N\) is the total number of samples.

  • Accuracy = (Number of correct predictions) / (Total samples)

  • Macro Precision = average of per-class precision

    • \[\text{MacroPrecision} = \frac{1}{C} \sum_{i=1}^{C} \frac{TP_i}{TP_i + FP_i}\]

  • Macro Recall = average of per-class recall

    • \[\text{MacroRecall} = \frac{1}{C} \sum_{i=1}^{C} \frac{TP_i}{TP_i + FN_i}\]

  • Macro F1 = average of per-class F1

    • \[\text{MacroF1} = \frac{1}{C} \sum_{i=1}^{C} \text{F1}_i\]

  • Weighted F1

    • \[\text{WeightedF1} = \sum_{i=1}^{C} \frac{n_i}{N} \cdot \text{F1}_i\]

Multi-label Classification (multiple labels/sample)

Let \(Y_i\) be the true labels and \(\hat{Y}_i\) the predicted labels for instance i.

  • Subset Accuracy (Exact Match)

    \[\text{Accuracy} = \frac{1}{N} \sum_{i=1}^{N} \mathbf{1}(Y_i = \hat{Y}_i)\]

  • Micro Precision / Recall / F1 (aggregate TP/FP/FN across all labels)

    • \[\text{MicroPrecision} = \frac{\sum_{l} TP_l}{\sum_{l} (TP_l + FP_l)}\]
    • \[\text{MicroRecall} = \frac{\sum_{l} TP_l}{\sum_{l} (TP_l + FN_l)}\]
    • \[\text{MicroF1} = 2 \cdot \frac{\text{MicroPrecision} \cdot \text{MicroRecall}}{\text{MicroPrecision} + \text{MicroRecall}}\]

  • Macro F1 (average across labels)

    • \[\text{MacroF1} = \frac{1}{L} \sum_{l=1}^{L} \text{F1}_l\]

Multi-task Binary Classification (1 binary label per task)

Let there be T tasks. Each task is a binary classification problem.

  • For each task, compute:

    • \(Accuracy\_t = (TP + TN) / (TP + FP + TN + FN)\)

    • \(AUC\_t\), \(F1\_t\) (as per binary classification)

  • MacroAUC / MacroF1 (across tasks)

    \[\text{MacroF1}_{\text{tasks}} = \frac{1}{T} \sum_{t=1}^{T} \text{F1}_t\]

Slice-based Performance Evaluation

How to choose slices for evaluation?

  • Numerical features: Quantile-based bins (e.g., age groups).

  • Categorical features: Stratify by value distribution.

  • Temporal features: Time-based slices (e.g., recent vs past).

  • Edge cases: Identify rare but critical scenarios.

When is a model ready for production?

  • Stable performance across test & validation sets.

  • Performs better than baseline (existing model or heuristic).

  • Low failure rate in stress tests (edge cases, adversarial inputs).

Model Evaluation Beyond AUC

  • Calibration: Platt scaling, isotonic regression.

  • Expected Calibration Error (ECE): Ensuring confidence scores are well-calibrated.

  • Robustness Testing: Adversarial robustness, stress testing with synthetic data.

Online Evaluation