SMOTE

An oversampling method for class imbalanceclass imbalance, n.: When one category is far rarer than the others in the data — say, one fraud case per thousand transactions — so a model can look accurate while missing nearly every rare case that matters. that creates synthetic minority-class training examples instead of duplicating existing ones. For each minority example it finds its k nearest minority-class neighbors (k = 5 by default), picks one at random and generates a new point at a random position on the line segment between them in feature space; repeating this until the desired class ratio is reached enlarges and smooths the minority region, encouraging broader decision boundariesdecision boundary, n.: The dividing line, or surface, where a classifier switches from predicting one category to another. Simple models draw straight lines; flexible ones draw curves. than plain duplication. It is usually combined with undersampling of the majority class and must be applied only to training data: resampling before splitting leaks synthetic near-copies of validation examples into training. Variants include Borderline-SMOTE, which oversamples near the class boundary; ADASYN, which focuses on hard regions; SMOTE-NC for mixed categorical and numeric features; and SMOTE-Tomek and SMOTE-ENN, which also remove overlapping points. Gains are inconsistent with strong learners such as gradient boostinggradient boosting, n.: A method that builds a strong predictor by adding small decision trees one at a time, each trained to correct the errors that remain after all the trees before it., and interpolation suits images, text or one-hot features poorly.

We applied SMOTE inside each cross-validation fold, never before the split, to avoid leakage.