Synthetic Minority Over-sampling Technique (SMOTE)

Synthetic minority over-sampling technique (SMOTE) is a method for fixing imbalanced datasets — datasets where one class is far rarer than the other. Fraud, churn, and conversion are classic examples: the event you care about most might be one in a hundred or one in a thousand cases. Introduced by Nitesh Chawla and colleagues in 2002, SMOTE balances the data not by copying the rare cases but by creating new, synthetic ones. For each minority-class example, it finds the nearest similar minority examples and generates a new point somewhere along the line between them — a plausible blend of two real cases rather than a duplicate. Repeat this until the rare class is well represented, and a classifier trained on the result has enough minority examples to actually learn the pattern instead of glossing over it.

The problem SMOTE solves is real and common. When ninety-nine percent of cases are one class, a model can hit ninety-nine percent accuracy by always predicting the majority and never catching the rare event you built it to catch. That model is useless precisely where it matters. Simply duplicating the minority examples helps a little but tends to cause overfitting, because the model memorizes the same few points. SMOTE's insight is to synthesize new examples in the gaps between real ones, expanding the minority region so the classifier learns a fuller picture of what the rare class looks like. It has become one of the most cited techniques in imbalanced learning and spawned dozens of variants such as Borderline-SMOTE.

SMOTE is one of several tools for class imbalance, and the honest framing is that it is a strong default, not the only answer. The simplest alternatives are random over-sampling (duplicate minority cases — risks overfitting) and random under-sampling (drop majority cases — throws away data). SMOTE sits between them by inventing new minority cases rather than copying or discarding. Other routes do not touch the data at all: you can apply class weights so the algorithm penalizes minority errors more heavily, or move the decision threshold so the model flags positives more readily. Often the best results come from combining approaches — for example, SMOTE plus a cleaning step that removes noisy synthetic points, or SMOTE plus class weighting.

SMOTE has real limits worth respecting. Because it interpolates between existing points, it can create unrealistic examples when classes overlap, effectively inventing minority cases inside majority territory and confusing the model. It works on numeric features and needs care with categorical ones (variants like SMOTE-NC exist for mixed data). The cardinal rule is to apply SMOTE only to the training data, after splitting off the validation and test sets — synthesizing before the split leaks information and produces flattering, dishonest scores. And SMOTE balances the training distribution, not the real world, so you must still evaluate on the true imbalanced distribution using metrics suited to imbalance, such as precision, recall, and the area under the precision-recall curve, rather than raw accuracy.

Using SMOTE well begins with the split: separate your test and validation data first, then apply SMOTE only to the training fold. Synthesizing before splitting leaks synthetic neighbors across the boundary and inflates your scores, which is the single most common SMOTE mistake. Decide how much to over-sample — full balance is not always best; sometimes lifting the minority to a moderate share works better — and consider a smarter variant when the basic method struggles, such as Borderline-SMOTE for cases near the decision boundary. Then judge the result on the real, imbalanced distribution with metrics that reward catching the rare class, because accuracy will mislead you on imbalanced data.

Mind the failure modes. SMOTE can manufacture noise when classes overlap, so inspect or clean the synthetic points and compare against simpler baselines like class weighting before assuming it helped. Do not treat balancing as a guaranteed win — sometimes adjusting the decision threshold or the loss function does the job with less risk. Remember that SMOTE addresses imbalance, not other data problems like leakage, label noise, or too few real minority examples to interpolate from meaningfully; if you have only a handful of genuine rare cases, no amount of synthesis conjures real signal. Used with a clean split and honest evaluation, SMOTE is a reliable way to make a classifier finally pay attention to the rare event that matters.

Synthetic minority over-sampling technique (SMOTE), introduced by Chawla et al. in 2002, balances imbalanced data by synthesizing minority-class examples, and is one of the most cited methods in imbalanced learning.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where synthetic minority over-sampling technique (smote) is a core concern: