When Rebalancing Backfires: Fairness, Bias, and the Ethics of Resampling Minority Classes

Every imbalanced dataset is a quiet piece of testimony about the world. Fraud is rare. Default is rare. The disease the model is meant to catch is, mercifully, rare. When we rebalance that data, we are not merely adjusting a ratio — we are deciding to overrule the world’s own count of itself, and to do it in the name of people who were already underrepresented. A decision that consequential deserves more scrutiny than the single line of library code it usually takes.

The Case for Forcing the Model to Look

Begin where the engineers begin, because their reasoning is genuinely sound. A model trained to maximize overall accuracy on a lopsided dataset learns a lazy and seductive trick: predict the majority every time. If only a sliver of transactions are fraudulent, a classifier that calls everything legitimate will be right almost always, and useless exactly when it matters. This is the problem of Class Imbalance, and the rare class — the fraud, the tumor, the loan that will default — is precisely the thing we built the model to find.

So we intervene, and the toolkit is well established. Oversampling duplicates the minority; Undersampling thins the majority; and synthetic methods manufacture new minority examples by interpolating between real ones. SMOTE, introduced by Chawla and colleagues in 2002, is the canonical synthetic approach; ADASYN, from He and colleagues in 2008, refines it by weighting synthesis toward the minority points hardest to learn. If editing the data feels too invasive, you can change the loss instead — Class Weighting, Cost Sensitive Learning, or Focal Loss, which Lin and colleagues proposed in 2017 to down-weight the easy examples a model has already mastered. Libraries such as Imbalanced Learn compress all of this into a single line, and metrics like Balanced Accuracy — the mean of per-class recall — reward the model for finally noticing the rare case. Seen this way, rebalancing looks like an act of justice: we are taking the side of the underrepresented. Who could object to that?

The Flaw Hiding Inside “Balance”

The crack appears the moment you ask what, exactly, balancing optimizes. It optimizes a number — recall on the minority class, or balanced accuracy — and a number is a narrow, literal thing. It is not fairness. It is not justice. It is a count of how often the model agrees with a label.

Even on its own terms, the technical case is shakier than the reflex suggests. A 2024 study on SMOTE and its variants found that for ordinary tabular data handled by strong modern learners — gradient-boosted trees, a tuned random forest, even plain logistic regression — skipping rebalancing entirely is often just as good, with a clear benefit confined to cases of extreme imbalance. The same work observes that synthetic oversampling tends to copy minority points and smear artifacts along the decision boundary, inventing density where the world offered none. So the first uncomfortable possibility is this: we may be paying a moral price for a technical gain that, on the models most teams actually use, frequently does not exist.

When the Remedy Becomes the Carrier

Now turn the same evidence around. If synthetic minority points are interpolations between real minority individuals, then every manufactured example is a statistical composite of real people — and the rare class, in the settings that matter most, is made of the most vulnerable people in the dataset.

Here the question the engineers rarely ask becomes unavoidable: can fixing class imbalance introduce unfair algorithmic bias? The evidence says it can. Work on fairness-aware resampling, including the FairSMOTE study from 2021, documents that conventional rebalancing optimizes accuracy rather than fairness, and that measured bias can actually increase after class-balancing in the datasets studied, lowering fairness across protected subgroups rather than lifting it. The intervention sold as protection for the underrepresented can deepen the disadvantage of the most underrepresented people inside that very class. Same tool, same data, opposite conclusion — and the fix quietly becomes the mechanism of harm.

The Question Was Never Whether the Data Is Balanced

Thesis: Resampling a minority class is a moral act wearing the costume of a preprocessing step, and treating it as neutral data cleaning is precisely how bias gets laundered into systems that believe they have already dealt with it.

The danger was never SMOTE, or ADASYN, or any particular function call. The danger is the sentence that follows them: “We balanced the data.” That sentence closes a conversation that should have stayed open. A team that has rebalanced feels it has discharged its obligation to fairness; the box is checked, the ratio is even, the audit moves on. But balancing the classes answers a question about arithmetic, not about who the model will fail and at what cost. A system that has been made to look fair, and is trusted because it looks fair, is more dangerous than one that wears its imbalance honestly.

The People Inside the Minority Class

Follow the accounting the optimistic case uses, and notice who never appears in it. The ledger records improved recall on the minority class as the victory. It does not record the minority individual at all — except as a row to be copied, interpolated, and balanced into submission.

There are real ethical risks to resampling data in medical and credit decisions, and they fall hardest on exactly these people. A 2025 study on privacy leakage in synthetic minority oversampling shows that the manufactured samples can expose private information about the real individuals they were interpolated from. In a fraud or default model, the minority is the borrower already on the margin. In a clinical model, it is the patient with the rare condition. These are the people least able to absorb a re-identification, and they are handed two bills at once: the privacy risk of having their data synthesized, and the false assurance that a balanced system has treated them fairly. The minority pays twice — once in exposure, once in the misplaced confidence of everyone downstream. Did anyone think to ask them?

Where This Argument Could Break

I should name the conditions under which I would withdraw the charge. If rebalancing were routinely paired with subgroup fairness audits and explicit privacy safeguards, the objection would lose most of its force — the fault would lie in the unexamined default, not the technique. At genuinely extreme imbalance, where the alternative is a model that ignores the minority completely, careful resampling can be the more humane choice rather than the less. And if newer synthesis methods can be shown to preserve subgroup structure without leaking the individuals they draw from, the privacy and fairness objections shrink toward footnotes. The argument here is against complacency, not against the tools — and complacency is a habit, which means it can be unlearned.

The Question That Remains

Balance is the easiest kind of fairness to fake, because it is the only kind that fits in a single line of code. If a model can be made to look just by editing the data it learns from, then the real question is not whether we balanced the classes, but whether we ever intended to be fair to the people inside them — or only to the metric that stands in for them.

Ethically, Alan.