Synthetic Minority Data Often Redundant or Invalid

For two decades, the standard approach to address class-imbalanced learning problems has involved fabricating synthetic minority examples. The conventional method for validating these synthetic points has been to score them against the very data used to generate them, a check that inherently cannot fail. This research introduces a de-biased validity test that fundamentally re-evaluates this practice. The new test defines validity as a population quantity: the probability that a synthetic point genuinely belongs to the minority class. It employs a consistent estimator that scores synthetic points against withheld real data, providing a more objective measure. The study reveals that the classical test significantly underestimates true invalidity in a vast majority of cases (96-99% of method-by-imbalance-ratio cells), whereas the de-biased estimator accurately tracks it. Crucially, the research proves that validity is an intrinsic property of the data itself, not the generation method. Class overlap establishes an "invalidity floor" that no faithful generator can escape, implying that oversampling is redundant when classes are separable and invalid when they overlap. Across 91 methods, three classifiers, and diverse datasets from medicine and finance, the study found that gains from synthetic data over the best trivial baseline were negligible (median F1 score below 0.01) and often damaged calibration. The audit is released as a pip-installable test, shifting the burden of proof onto synthetic minority data methods to demonstrate both validity and information gain on the specific data at hand.