讨论： Handling class imbalance: metrics first, then weights, thresholds and resampling inside the pipeline

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

暂无译文，显示原文。 原文

The order 'class weights, then threshold, then resampling' puts the intervention with the fewest side effects last. For a probabilistic classifier, imbalance by itself does not damage the ranking; what a naive setup gets wrong is the 0.5 cut-off, and moving the threshold to the cost-derived operating point fixes the confusion matrix without touching the model or its probabilities. Weights and resampling change the fitted probabilities: the model then overstates the minority class, which the article's last bullet admits by requiring calibration afterwards, and that calibration costs a further held-out set. A simulation study in JAMIA (van den Goorbergh and colleagues, 2022) reported for logistic regression that random undersampling, oversampling and SMOTE produced strongly miscalibrated models without higher ROC AUC than no correction, and that the improved balance between sensitivity and specificity was obtained equally by shifting the threshold. So for models that output probabilities the order should be: fix evaluation, tune the threshold, and reach for weights or resampling only when the ranking itself is poor and a cheaper fix (more minority data, better features) is unavailable. Weights keep a real role for loss functions that ignore the minority class entirely, such as a tree with a minimum leaf size larger than the minority count.

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-17

暂无译文，显示原文。 原文

'Put the sampler inside the pipeline' needs one specific class: scikit-learn's `Pipeline` has no notion of a step that changes the number of rows, so a sampler placed in it fails. imbalanced-learn ships `imblearn.pipeline.Pipeline`, whose steps may implement `fit_resample`; those steps run during `fit` only and are skipped during `predict`, `transform` and `score`, which is exactly the behaviour the article asks for (validation folds keep their natural ratio). The `class_weight="balanced"` formula the API documents is `n_samples / (n_classes * np.bincount(y))`, so the weights are set by the training fold's frequencies and change with every resample; and `SMOTENC` is the SMOTE variant for mixed numeric and categorical columns, since plain SMOTE interpolates categorical codes.