What actually matters in synthetic oversampling? A factorial decomposition of SMOTE-family methods and an evidence-based configuration

Abstract Around a hundred variants of the synthetic minority oversampling technique (SMOTE) have been proposed, each changing some part of the original algorithm, yet it is rarely known which part is responsible for a reported gain. We decompose SMOTE-family oversamplers into five interchangeable components (seed allocation, partner selection, generator, filter and budget) and study them in a full factorial design complemented by level sweeps, reference strategies and eleven published methods. Fifty-four imbalanced data sets, six learners, repeated 5×2 cross-validation and an inner-validated decision threshold are combined with Bayesian hierarchical models that separate component effects from data-set and learner heterogeneity and assess practical equivalence with a region of practical equivalence (ROPE). Partner selection was practically inert in both learner groups and for every neighbourhood size. Allocation, filtering and budget mainly shifted the operating point between recall and precision, a shift that a tuned threshold removed. Edited-nearest-neighbour cleaning was consistently harmful, and difficulty-based (ADASYN-type) allocation slightly so. The largest gains came from volume-filling generators combined with tree ensembles, about two points of PR-AUC and tuned MCC, larger on average in data sets whose minority examples overlap the majority class. Borderline-SMOTE, ADASYN, SMOTE-ENN and G-SMOTE behaved like single components. A configuration derived from these results, safety-allocated local Gaussian oversampling, retained a smaller advantage over SMOTE on 57 hold-out data sets, credible in PR-AUC for both learner groups. Component terms explained less than one percent of the variance of PR-AUC and tuned MCC, far less than the data set.