Does SMOTE just duplicate minority rows?

Basic SMOTE interpolates new feature vectors; ordinary random oversampling may duplicate rows.

Does it work directly on any categorical data?

No; suitable variants or other approaches are needed.

Should the test set be oversampled too?

Usually not; it should represent the evaluation target.

AI Glossary / Statistics and probability / Intermediate

Synthetic minority oversampling technique

Definition

Synthetic minority oversampling technique, commonly called SMOTE, is a training-data resampling method that creates additional minority-class examples by interpolating between existing minority examples and nearby minority neighbors. It is used in imbalanced classification. It changes the training sample's representation, but does not create independently observed evidence or guarantee better model performance.

What it means

When one class is much less frequent than another, a model may learn patterns that favor the majority class. SMOTE offers one way to modify the training data by adding points in minority-class neighborhoods instead of simply repeating the same records.
The generated examples inherit the minority label under the method's assumptions. They are mathematical constructions in feature space, not newly verified real cases. Whether their interpolation is meaningful depends on the features, local class structure, and evaluation objective.

How it works

Separate training and evaluation data before resampling. In cross-validation, fit and apply SMOTE only within each training fold. Resampling the entire dataset first can let information from validation cases influence synthetic training points and distort the evaluation distribution.
Choose an appropriate feature representation and distance metric. Basic SMOTE assumes numerical feature vectors where interpolation and nearest-neighbor distances make sense. Feature scaling can strongly influence which examples count as neighbors.
For a selected minority example, identify a specified number of nearby minority examples. Choose a neighbor and draw an interpolation factor between zero and one. Generate a point along the line segment joining the two feature vectors.
For seed x, neighbor z, and factor a, the synthetic point is x + a(z-x). Its coordinates are interpolated together using the chosen factor. Repeat as needed to reach the selected resampling target; complete class balance is not mandatory.
Train the classifier on the resulting training data. Tune the resampling amount, neighbor settings, model, and decision threshold using a leakage-safe evaluation process. Compare against simpler alternatives such as class weighting or threshold adjustment.
Check whether the synthetic points remain plausible. Interpolation can cross class boundaries, bridge separate minority clusters, or amplify mislabeled examples. In high-dimensional spaces, distance may be less informative and local neighborhoods can be unreliable.
Categorical or mixed features require suitable variants or alternative methods. Interpolating category codes can create meaningless values. Domain constraints may also make ordinary numerical interpolation invalid even when every column is numeric.
Evaluate on held-out data representative of intended use, without artificially balancing it to match the training resample. Examine class-specific performance, precision-recall behavior, calibration, and the costs of errors rather than relying on accuracy alone.
Because resampling changes class proportions, raw probability estimates may need calibration or other appropriate adjustment for deployment. A larger resampled row count does not imply that the effective amount of independent information has increased by the same amount.

Three practical examples

Defect classification: explore whether minority-class interpolation improves detection on untouched inspection data.
Support routing: compare resampling with class weighting when rare issue categories are poorly recognized.
AI model development: include SMOTE in a training pipeline while preserving realistic class prevalence in evaluation.

Think of it this way

Think of filling gaps between nearby examples on a map. The method draws additional points between known minority locations. That can make the region easier for a learner to recognize, but a straight line between two valid locations does not guarantee every point on it represents a valid real case.

Worked example

Consider an original two-feature example after a suitable feature transformation. A minority seed has coordinates x = (2,4), and a selected minority neighbor has coordinates z = (6,10). Choose interpolation factor a = 0.25.
The difference z-x is (4,6). The synthetic point is (2,4) + 0.25(4,6) = (3,5.5). It lies one quarter of the way from the seed to the neighbor and is assigned the minority label by the resampling method.
Suppose the original training set has 900 majority and 100 minority examples. Adding 200 synthetic minority points produces 900 majority and 300 minority rows, for 1,200 rows total. The minority share changes from 100/1,000 = 10% to 300/1,200 = 25%.
This illustrates a chosen partial resampling target, not a requirement to balance classes. The original number of independently observed minority cases remains 100. A separate test set should retain a distribution appropriate to deployment rather than being resampled merely to resemble this training set.

Illustrative case study

Fictional case: Pine Inspection evaluates a classifier for rare manufacturing defects. An initial experiment applies SMOTE before splitting the data and reports a large improvement. A reviewer identifies leakage and rebuilds the workflow so resampling happens only inside training folds. The revised evaluation shows a smaller gain and some loss of precision. The team compares error costs, checks generated feature plausibility, and decides whether the improvement is useful under actual defect prevalence.

Common misconceptions and mistakes

Do not resample before separating evaluation data. Do not treat generated examples as independently collected observations. SMOTE does not guarantee realistic synthetic cases or eliminate class overlap. Numerically encoded categories are not automatically suitable for interpolation. A balanced training set does not require a balanced test set. Better recall alone may be insufficient if false positives become operationally costly.

Business use

Use SMOTE as a candidate training intervention for imbalanced classification, with a clear objective and controlled comparisons. Document where resampling occurs, how neighborhoods are defined, and what class ratio is targeted. Decide using realistic held-out performance, probability quality, and operational consequences rather than the size of the resampled dataset.

Check your understanding

What synthetic point is generated in the example?

Answer: (3,5.5).

What is the minority share after adding 200 points?

Answer: Twenty-five percent.

How many original minority observations were collected?

Answer: One hundred, despite the larger resampled count.

Related terms

Oversampling
Nearest neighbor
Data leakage
Probability calibration
Class imbalance
nearest neighbors
interpolation
class weighting
precision-recall

Learn next

Compare SMOTE with random oversampling and class-weighted learning in a properly nested training pipeline. Then examine variants for mixed features and methods that address noisy or overlapping minority neighborhoods.

Sources

Chawla and colleagues, SMOTE: Synthetic Minority Over-sampling Technique: https://www.cs.cmu.edu/afs/cs/project/jair/pub/volume16/chawla02a.pdf imbalanced-learn, Common pitfalls and recommended practices: https://imbalanced-learn.org/stable/common_pitfalls.html imbalanced-learn, Over-sampling, mathematical formulation: https://imbalanced-learn.org/stable/over_sampling.html The interpolation and class-count example is original.

Previous: Synthetic measure Next: System of record