Under-Sampling Majority Class

Under-Sampling Majority Class is the class imbalance technique that reduces the majority class by removing samples — creating a balanced training set by discarding excess majority examples, trading off majority class information for balanced training.

Under-Sampling Methods

• Random Under-Sampling: Randomly remove majority samples — simple but loses information.
• NearMiss: Select majority samples close to minority decision boundaries — keep the informative ones.
• Tomek Links: Remove majority samples that form Tomek links (closest pairs of opposite classes) — clean decision boundary.
• Cluster Centroids: Cluster majority samples and keep only centroids — preserves distribution structure.

Why It Matters

• Fast Training: Smaller balanced dataset trains much faster than the full imbalanced dataset.
• Information Loss: The main drawback — discarding majority samples loses potentially useful information.
• Complementary: Often combined with over-sampling (SMOTE + Tomek Links) for better results.

Under-Sampling is trimming the majority — reducing dominant class samples to create a balanced training set at the cost of some information loss.