undersampling

Undersampling is the deliberate reduction of the sampling rate or the number of samples in a dataset. In signal processing, undersampling refers to sampling a continuous signal at a rate below the Nyquist rate, which can lead to aliasing and distortion unless mitigated by pre-sampling filtering. In machine learning and data mining, undersampling denotes techniques for balancing imbalanced data by reducing the size of the majority class.

In signal processing, undersampling must be preceded by an anti-aliasing filter that removes frequency components above half the target sampling rate. Practical uses include bandwidth reduction, data compression, or streaming where lower rates save bandwidth or storage. The downside is information loss and potential distortion if the filter does not fully remove high-frequency content.

In machine learning, undersampling methods aim to balance class distributions. Random undersampling randomly removes samples from the majority class until a desired balance is reached, risking loss of informative data. More sophisticated approaches—such as informed or cluster-based undersampling—select representative samples to preserve information while reducing size. Techniques like removing borderline samples (for example, using Tomek links or edited nearest neighbors) are often used. Undersampling is frequently combined with oversampling of the minority class in hybrid approaches to improve classifier performance. Evaluation typically uses metrics sensitive to class imbalance, such as precision, recall, F1 score, and ROC-AUC.

Overall, undersampling can reduce data imbalance and processing requirements but may also discard useful information and increase variance, necessitating careful design and validation.