NearMiss

Selects majority class samples whose average distance to the three closest minority class instances is the smallest.

• Mechanism: For each majority sample, compute the mean distance to its k nearest minority neighbors. Keep the m majority samples with the lowest mean distances.
• Effect: Retains majority examples that are closest to the minority class, effectively focusing the classifier on the most ambiguous region near the decision boundary.
• Risk: Can amplify noise if the minority class contains outliers, as the algorithm will aggressively select majority points near those outliers.
• Use Case: When the minority class is compact and well-defined, and you want the model to learn a tight boundary.

Selects majority class samples whose average distance to the three farthest minority class instances is the smallest.

• Mechanism: For each majority sample, compute the mean distance to its k farthest minority neighbors. Keep the m majority samples with the lowest mean distances.
• Effect: Retains majority examples that are closest to the periphery of the minority class distribution, providing a broader, more diverse selection of majority points around the minority class.
• Advantage: Less sensitive to minority class outliers compared to NearMiss-1, as it considers the farthest minority neighbors rather than the nearest.
• Use Case: When the minority class has scattered sub-clusters and you need wider coverage around each minority instance.

A two-stage algorithm that, for each minority class instance, selects a fixed number of the closest majority class neighbors.

• Mechanism: For each minority sample, keep its m nearest majority neighbors. This directly balances the dataset by ensuring each minority example is surrounded by a controlled number of majority examples.
• Effect: Guarantees that every minority instance is retained and paired with a local neighborhood of majority points, preserving the local density structure around each rare event.
• Contrast: Unlike NearMiss-1 and NearMiss-2, which select a global count of majority samples, NearMiss-3 makes a per-minority-instance selection, inherently linking the sampling to the minority class distribution.
• Use Case: When preserving the local topology around every fraud example is critical for the downstream classifier.

The choice of distance metric fundamentally alters which majority instances are selected by any NearMiss variant.

• Euclidean Distance: Default metric. Sensitive to feature scaling; requires normalization. Works well for continuous, normally distributed features.
• Manhattan Distance: Less sensitive to outliers in high-dimensional spaces. Preferable when features are sparse or have heavy-tailed distributions.
• Mahalanobis Distance: Accounts for feature covariance, making it scale-invariant. Critical when transaction features are highly correlated.
• Cosine Similarity: Focuses on angular separation rather than magnitude. Useful when the direction of a transaction vector matters more than its absolute value.
• Best Practice: Always standardize features before applying any NearMiss variant, and experiment with multiple distance metrics on a held-out validation set.

NearMiss algorithms rely on distance-based heuristics that make implicit assumptions about the data geometry, which can fail in practice.

• Curse of Dimensionality: In high-dimensional feature spaces common in fraud detection, pairwise distances become less meaningful. NearMiss may select uninformative majority samples.
• Noisy Minority Instances: NearMiss-1 is particularly vulnerable to selecting majority points near mislabeled or anomalous minority examples, degrading the decision boundary.
• No Statistical Guarantee: Unlike cost-sensitive learning or ensemble methods, NearMiss provides no theoretical guarantee on preserving the true underlying distribution.
• Mitigation: Combine NearMiss with Tomek Links or Edited Nearest Neighbors as a post-cleaning step to remove ambiguous majority instances retained near the boundary.

NearMiss can be embedded within bagging and boosting frameworks to recover information lost during under-sampling.

• Bagging + NearMiss: Train multiple base classifiers, each on a different NearMiss-1 under-sampled subset of the majority class. Aggregate predictions via majority voting to reduce variance.
• Boosting + NearMiss: At each boosting iteration, apply NearMiss-2 to the majority class before training the next weak learner. This forces subsequent learners to focus on increasingly difficult boundary regions.
• Hybrid Pipeline: Apply NearMiss-3 for initial balancing, then use SMOTE to over-sample the minority class, followed by a gradient-boosted tree model with scale_pos_weight tuning.
• Library Support: The imbalanced-learn library provides NearMiss as a scikit-learn compatible sampler, enabling seamless integration into Pipeline objects with RandomUnderSampler or ensemble wrappers like BalancedBaggingClassifier.