Does SMOTE duplicate existing rows?

Basic SMOTE interpolates between examples, unlike simple random oversampling that repeats selected rows.

Can it handle multiple classes?

Implementations can oversample selected classes in multiclass settings; the strategy must be specified.

Does synthetic mean anonymous?

No. SMOTE is not a privacy guarantee or a substitute for assessing disclosure risk.

AI Glossary / Data science and analytics / Intermediate

SMOTE

Definition

SMOTE stands for Synthetic Minority Over-sampling Technique. It creates additional training examples for an underrepresented class by interpolating between existing examples of that class. The original method was introduced by Chawla, Bowyer, Hall and Kegelmeyer in 2002 for learning from imbalanced classification data.

What it means

A classifier may learn little about a rare class when most training records belong to another class. For example, a machine-condition dataset might contain many normal observations and relatively few labeled faults. SMOTE changes the training data presented to the learner by adding synthetic minority-class feature vectors.

These vectors are constructed from observed examples; they are not newly measured faults or independent evidence about the population. Their usefulness depends on whether interpolation in the chosen feature space produces plausible class members.

SMOTE is a training technique, not a classifier, a guarantee of improved accuracy or a requirement whenever class counts differ. Compare it with appropriate baselines and assess performance on real held-out observations.

How it works

In basic SMOTE, choose a minority-class example and one of its nearby minority-class neighbors. Generate a point partway along the line between their feature vectors. In notation, x_new = x + lambda(y − x), where lambda lies between zero and one.

The neighbor count, distance measure, feature scaling and amount of oversampling affect the generated points. Numeric features on incompatible scales can distort what “nearby” means. Ordinary interpolation is unsuitable for arbitrary category codes; specialized variants address mixed or categorical data differently.

Split evaluation data before learning preprocessing or generating synthetic examples. During cross-validation, fit resampling within each training fold, rather than oversampling once before making folds. Keep evaluation observations real and aligned with the intended deployment distribution.

Finally, compare models using measures suited to the decision, including minority-class recall, precision, false-alarm burden and probability calibration where relevant. Balancing counts alone is not the objective.

Three practical examples

A manufacturing team tests whether SMOTE helps recognize an uncommon fault type from suitable continuous sensor features, using later real observations for evaluation.
An AI service compares oversampling with class-weighted training when categorizing a rarely occurring support-request type.
A small research project uses synthetic points to explore a classifier's sensitivity to imbalance, while clearly distinguishing generated records from collected measurements.

Think of it this way

Think of drawing extra practice locations along a path between two known landmarks. The added locations make the learner practice the space between observed examples rather than repeatedly visiting only the landmarks.

The analogy also exposes the limitation: a straight path may cross terrain that does not belong to the same region. Two minority examples can surround majority-class territory or an impossible feature combination. Interpolation is an assumption about useful training geometry, not proof of real-world validity.

Worked example

Generating one illustrative point

Suppose two nearby minority examples have two already prepared numerical features: x = (2, 4) and y = (6, 10). Choose lambda = 0.25 for this teaching calculation.

The difference y − x is (4, 6). Multiplying by 0.25 gives (1, 1.5), and adding it to x produces x_new = (3, 5.5). The generated point receives the minority-class training label.

This does not show that a real observation at (3, 5.5) would belong to that class. The calculation only explains the construction. If the second feature were a category identifier, the value 5.5 could be meaningless.

If a training set contains 900 majority and 100 minority records, adding 200 synthetic minority records changes the training counts to 900 and 300. It does not change the estimated real-world event frequency to 25%, and it adds no independently observed minority cases.

Illustrative case study

Fictional case: Summit Sensors evaluates a rare-fault detector. An initial experiment applies SMOTE to the entire dataset and then randomly divides the resulting records into training and testing portions.

A reviewer notices that synthetic training points can depend on observations assigned to the test portion. The reported evaluation therefore no longer provides a clean separation from model development.

The team rebuilds the experiment using a time-aware holdout and resampling only inside training folds. It compares an unchanged baseline, class weighting and SMOTE, examining false alarms as well as detected faults. No winner is assumed. The scenario illustrates how a plausible improvement claim must survive an evaluation design that excludes information leakage.

Common misconceptions and mistakes

Synthetic records are equivalent to additional independent observations. They derive from existing examples and do not supply new measurements.
Equal class counts guarantee a good model. The chosen decision threshold, errors and data geometry still matter.
SMOTE should be applied to the test set too. Doing so changes the evaluation question and can introduce leakage.
Any numbers can be interpolated meaningfully. Encoded categories and constrained feature combinations require special care.

Business use

SMOTE can be worth evaluating when a consequential class has limited training representation and the features support meaningful neighborhoods. It may also magnify noise or create points across class boundaries, so inspect the original labels and generated feature combinations.

Business evaluation should translate prediction errors into workload and consequences. A detector that catches more rare events but produces an unmanageable review queue may not improve the process. Preserve the real class distribution in decision-focused testing, and assess probabilities separately because changing training prevalence can affect their interpretation.

Check your understanding

A team uses test-set examples as neighbors when generating synthetic training points, but never directly trains on the original test rows. Is the test set still independent of development?

Answer: No. Information from those observations influenced the training data. Generate synthetic examples using only the relevant training partition and evaluate on untouched real observations.

Related terms

Oversampling
Undersampling
Nearest neighbor
Data leakage
Precision
Recall
Calibration
Class imbalance
nearest neighbours
class weighting

Learn next

Read class imbalance and data leakage. Design a comparison between an unchanged baseline, class weighting and SMOTE, specifying the split, training-only preprocessing, relevant error measures and operational decision threshold before judging results.

Data leakage

Sources

Chawla, Bowyer, Hall and Kegelmeyer, SMOTE: Synthetic Minority Over-sampling Technique: https://arxiv.org/abs/1106.1813. imbalanced-learn, Over-sampling and mathematical construction: https://imbalanced-learn.org/stable/over_sampling.html. imbalanced-learn, Common pitfalls: https://imbalanced-learn.org/stable/common_pitfalls.html. Numerical examples and Summit Sensors are original illustrations, not experimental findings.

Previous: Smoothness term Next: Snapshot isolation