sampling/

Adds minority examples, keeping every majority row.

• RandomOverSampler: Duplicates minority rows, optionally jittered.
• SMOTESampler: Interpolates new points between neighbouring minority

• BorderlineSMOTESampler: SMOTE concentrated near the decision boundary,

• ADASYNSampler: Generates more where the minority is hardest to learn.
• SVMSMOTESampler / KMeansSMOTESampler: SMOTE guided by an SVM boundary

• ClusterOverSampler: Cluster-aware oversampling.

Drops majority examples. Faster to train, at the cost of discarding data.

• RandomUnderSampler: Removes majority rows at random.
• TomekLinksSampler: Removes majority rows that sit right against a

• ENNSampler: Edited Nearest Neighbours cleaning.
• CNNSampler: Condensed Nearest Neighbour condensing.
• NearMissSampler: Distance-based selection.
• HardnessThresholdSampler: Drops the rows a classifier finds easy.

• ClassBalanceSampler: Balance the class distribution directly.
• ReservoirSampler: Fixed-size random sample from a stream.

Resample the training split only. Synthetic or duplicated rows in a test set produce a score that cannot be reproduced on real data — and with oversampling, near-copies of training rows end up in test, which inflates it outright. Passing these as pipeline steps to train handles this: they are applied inside each fold, to the training half.

Accuracy stays misleading after resampling. Judge on the original distribution, with a metric that respects the minority class.