ADASYN

When one class is rare, a model can get lazy and mostly learn the majority class. ADASYN is a way to fight that by creating new minority-class training examples, with extra attention paid to the minority cases that are hardest to learn.

What it is doing

ADASYN stands for Adaptive Synthetic Sampling. It is an oversampling method for imbalanced classification. Like SMOTE, it generates synthetic minority examples instead of simply duplicating existing ones. The key difference is the word adaptive: ADASYN creates more synthetic data near minority points that sit in difficult regions, especially where many nearby examples belong to the majority class. The idea is that these are the places where the classifier needs the most help.

How it works in practice

For each minority example, ADASYN looks at its nearest neighbors and estimates how hard that example is to learn. Minority points surrounded by majority points get a higher weight, so they receive more synthetic samples. New samples are then created by interpolating between a minority point and nearby minority neighbors.

In fraud detection, rare fraudulent transactions near normal-looking purchases are exactly the tricky cases ADASYN emphasizes.
In disease diagnosis, borderline positive cases that resemble negatives can get more representation in training.
In Python, this is commonly used through imblearn.over_sampling.ADASYN.

Why it matters

ADASYN can improve a classifier’s ability to detect the minority class, which matters in tasks like churn prediction, credit default, or spam detection where missing rare cases is costly. But it is not automatically better than simpler methods: by focusing on hard regions, it can also amplify noise or outliers if the minority class contains mislabeled points. That is why it should be applied only to the training set, then judged with imbalance-aware metrics such as recall, precision, F1, or PR AUC, not plain accuracy.

ADASYN (Adaptive Synthetic Sampling) is an oversampling method for imbalanced classification that generates synthetic minority-class examples in feature space, allocating more new samples to minority points that are harder to learn (typically those surrounded by majority-class neighbors). It matters because it shifts training emphasis toward difficult regions, improving minority recall and decision boundary quality when class imbalance would otherwise bias supervised models toward the majority class.

Imagine you’re studying for a test using flashcards, but you only have a few cards for the hardest topics. A smart friend helps by making new practice cards that look like the tricky ones you already struggle with, so you get more practice where you need it most.

ADASYN (Adaptive Synthetic Sampling) does something similar for AI training data when one class is rare—like fraud cases compared to normal purchases. Instead of just copying rare examples, it creates new, realistic “made-up” examples, focusing more on the rare cases that are hardest to learn (the ones near confusing boundaries). This helps the model pay attention to the tough, underrepresented situations.