Does ADASYN create new real-world evidence?

No. It creates synthetic training points from existing observations and assumptions about their neighborhoods.

Is it identical to SMOTE?

No. Both use synthetic interpolation, but ADASYN adapts how much to generate around each seed using a local difficulty estimate.

Can it be applied directly to arbitrary text or category codes?

Ordinary feature-space interpolation is not automatically meaningful for those representations. Use a method appropriate to the data and validate the generated examples.

AI Glossary / Generative AI and language / Intermediate

Adaptive Synthetic Sampling

Definition

Adaptive Synthetic Sampling, usually abbreviated ADASYN, is an oversampling method for imbalanced classification. It generates synthetic minority-class examples, allocating more examples around minority observations whose local neighborhood contains a higher proportion of majority-class observations. It was introduced by He and colleagues in 2008.

What it means

A classifier trained on imbalanced data may give insufficient attention to a rare class that matters to the task. ADASYN changes the training dataset by adding constructed minority examples rather than merely duplicating every minority row equally.

Its adaptation concerns where to generate those examples. A minority observation surrounded by many majority observations is treated as more difficult, so it receives a larger share of the generation budget.

That neighborhood measure is a heuristic. A difficult region could contain a genuine class boundary, overlapping classes, an unusual case or a labeling error. More synthetic data does not automatically resolve those possibilities. ADASYN creates training points from existing information; it does not collect new independent evidence about the real population.

How it works

In the original binary-classification method, choose the desired amount of oversampling and a number of neighbors. For each minority example, inspect its nearest neighbors in the dataset and calculate the proportion belonging to the majority class.

Normalize these local difficulty values to allocate the total synthetic-example budget. Minority points with larger values receive more generated examples. If every difficulty value is zero, this normalization has no positive total; the implementation needs explicit handling rather than an invented allocation.

For generation, select a suitable minority neighbor and interpolate between its feature vector and the seed. A typical synthetic point is x + lambda times (neighbor - x), where lambda lies between zero and one. It receives the minority label.

Fit resampling only within the training portion, including inside cross-validation folds. Keep assessment data representative of the intended population. Distances depend on feature representation and scaling, and ordinary numerical interpolation may produce invalid categorical values or impossible feature combinations.

Three practical examples

A defect classifier's developers compare ADASYN with class weighting when rare defect examples are underrepresented in the fitting data.
A numerical support-request model experiments with additional minority examples near overlapping class regions, then evaluates on untouched requests.
A research team studies whether neighborhood-focused oversampling improves minority recall without creating an unacceptable rise in false alarms.

Think of it this way

Think of creating extra practice exercises around topics where a learner struggles, rather than repeating every topic equally. The allocation is targeted using a difficulty signal.

The analogy has limits: ADASYN constructs feature vectors by interpolation, not carefully authored exercises. If the difficulty signal reflects a mistaken label or an impossible example, extra practice generated around it can reinforce the problem.

Worked example

Allocating and constructing examples

Consider an invented allocation step with a budget of 16 synthetic examples. Three minority seeds have nonzero difficulty values 0.2, 0.6 and 0.8; all other seeds have zero. Their total is 1.6.

The normalized shares are 0.125, 0.375 and 0.5, giving allocations of 2, 6 and 8 examples. Real implementations need a rounding policy when allocations are not integers, so achieved totals can differ from a requested target.

For one generated point, suppose a seed is (2, 4), its chosen minority neighbor is (6, 8), and lambda is 0.25. Interpolation gives (2, 4) + 0.25 times (4, 4) = (3, 5).

This arithmetic does not establish that (3, 5) is physically possible or correctly labeled. The feature meaning and local class structure still need scrutiny.

Illustrative case study

A fictional component supplier, Willow Inspection, applies ADASYN to a numerical defect dataset. Its first experiment shows a large apparent improvement after the entire dataset was resampled and only then divided into training and test portions.

The analysts recognize that synthetic examples can carry information across that boundary and that the test distribution no longer represents ordinary incoming parts. They restart with the original data split first, resampling only the training portion.

They also investigate isolated minority observations before generating extra examples near them. Some are valid rare defects; others contain measurement errors. The fictional case demonstrates that correcting the evaluation design and source data is necessary before interpreting any performance gain.

Common misconceptions and mistakes

ADASYN is not simply random duplication, and it does not necessarily generate equal numbers around every minority example. It allocates according to local class composition.

Do not assume that full class balance is always the best target. The appropriate training strategy depends on the learner, error costs and operating population.

Do not treat synthetic rows as additional independent observed cases or use their count to inflate confidence in evaluation results. Oversampling can also alter probability behavior, so assess calibration and thresholds on suitable untouched data rather than assuming predicted probabilities retain their original interpretation.

Business use

Use ADASYN as a candidate training strategy, not a default cure for rare classes. Compare it with simpler baselines such as class weighting, threshold adjustment and other sampling methods under the same evaluation design.

Inspect minority examples and feature geometry before interpolation. Measure relevant precision-recall trade-offs and downstream review demand, not just accuracy on a balanced sample. Preserve the resampling configuration with the training record so the procedure can be reproduced and reassessed when the data change.

Check your understanding

Should ADASYN be run on all records before creating a test split so both portions contain balanced classes?

Answer: No. Split the original data first and resample only fitting data. Otherwise information can cross the evaluation boundary, and the test distribution may no longer represent the intended population.

Related terms

Oversampling
Minority class
SMOTE
Nearest neighbor
Data leakage
Class imbalance
nearest neighbours
interpolation
class weighting
precision and recall

Learn next

Calculate local majority-neighbor proportions for a small labeled dataset and inspect which seeds receive more synthetic examples. Compare the generated points with feature constraints. Then evaluate an entire resampling pipeline against an untouched, representative test set and a simple baseline.

Test set

Sources

Sources: He, Bai, Garcia and Li, ADASYN: Adaptive Synthetic Sampling Approach for Imbalanced Learning: https://sci2s.ugr.es/keel/pdf/algorithm/congreso/2008-He-ieee.pdf Imbalanced-learn, ADASYN implementation: https://imbalanced-learn.org/stable/references/generated/imblearn.over_sampling.ADASYN.html Imbalanced-learn, resampling leakage and evaluation pitfalls: https://imbalanced-learn.org/stable/common_pitfalls.html

Previous: Adaptive softmax Next: Adaptive utility