Notes
🔍

Random Under-Sampling

When one class is rare (fraud, disease, churn) and the other is everywhere, a model can get “good” accuracy by mostly predicting the common class. Random under-sampling is a simple way to stop that by shrinking the majority class in the training data.

What it is and how it works

Random under-sampling means you randomly remove examples from the majority class until the class counts reach a chosen balance (often 1:1, but not required). You keep all (or nearly all) minority examples and throw away a random subset of majority examples. Conceptually, it’s like editing a huge stack of “non-fraud” transactions down to a manageable, representative sample so the learner is forced to pay attention to fraud patterns.

Where it shows up in practice

  • Fraud detection: millions of legitimate transactions vs. a tiny number of fraud cases.
  • Medical screening: many healthy patients vs. few positives.
  • Customer churn: most customers stay; a minority churn.

In Python, you’ll commonly see this via imbalanced-learn’s RandomUnderSampler:

from imblearn.under_sampling import RandomUnderSampler
X_res, y_res = RandomUnderSampler(sampling_strategy=0.5, random_state=0).fit_resample(X, y)

Why it matters (and the trade-offs)

Under-sampling can dramatically improve a classifier’s ability to learn the minority class, speed up training, and reduce memory use. The cost is information loss: you might discard important majority patterns, which can hurt calibration and increase false positives. It works best when the majority class is very large and redundant, and it should be applied only to the training split (never the test set), with metrics like precision/recall, F1, or PR-AUC guiding whether the trade-off is worth it.

Random Under-Sampling is a resampling technique for imbalanced supervised classification that randomly removes examples from the majority class to reduce class skew and produce a more balanced training set. It matters because many learners and metrics are biased toward the majority class; under-sampling can improve minority-class recall and decision boundaries, but it risks discarding informative majority examples and increasing variance if too aggressive.

Imagine you’re trying to learn what a rare disease looks like, but your photo album has 9,900 pictures of healthy people and only 100 of the disease. If you study the whole album, you might “learn” that everyone is healthy, because that’s what you see most.

Random Under-Sampling fixes this by randomly throwing away some of the extra “healthy” examples so the training set is more balanced. In supervised learning, this can help a model pay attention to the smaller, important group (like fraud cases or spam) instead of being overwhelmed by the common cases. The trade-off is you might discard useful information when you remove data.