SMOTETomek

The Synthetic Minority Over-sampling Technique generates artificial minority class examples by interpolating between existing data points in feature space. Unlike simple duplication, SMOTE creates new, plausible transactions along the line segments connecting a minority instance to its k-nearest neighbors.

• Mechanism: Selects a random neighbor and creates a synthetic point at a random distance along the vector between the two
• Advantage: Expands the minority class decision region, forcing the classifier to generalize rather than memorize
• Limitation: Can amplify noise and create samples deep in majority class territory without cleaning

A data cleaning method that identifies pairs of minimally distanced nearest neighbors from opposite classes. When two instances form a Tomek Link, they sit directly on the class boundary, and the majority class instance is removed to create a cleaner separation.

• Identification: Finds all pairs (x_i, x_j) where x_i and x_j are nearest neighbors but belong to different classes
• Action: Removes the majority class instance from each pair
• Result: Eliminates overlapping regions where classifiers struggle, sharpening the decision boundary

An under-sampling technique that removes any majority class example whose predicted class label differs from its actual label, based on a vote from its k-nearest neighbors. ENN acts as a noise filter, discarding ambiguous or mislabeled instances near the boundary.

• Process: For each majority sample, check if at least k/2 neighbors agree with its class; if not, remove it
• Effect: Cleans both overlapping and noisy interior points
• Use in SMOTEENN: Applied after SMOTE over-sampling to remove any synthetic samples that land in majority territory

A combined resampling strategy that applies SMOTE to over-sample the minority class, then immediately applies Edited Nearest Neighbors to clean the resulting dataset. Unlike SMOTETomek, which only removes majority instances in Tomek Links, SMOTEENN can remove examples from both classes if they are misclassified by their neighbors.

• Aggressive Cleaning: Removes any instance whose predicted class doesn't match its actual class
• Trade-off: More thorough noise removal but can discard more minority samples than SMOTETomek
• Best For: Datasets with significant label noise or overlapping distributions

A learning paradigm that assigns higher misclassification costs to the minority class, penalizing the model more severely for false negatives than false positives. This approach addresses imbalance at the algorithm level rather than the data level.

• Implementation: Modify the loss function or use class weights in frameworks like XGBoost (scale_pos_weight) or LightGBM (is_unbalance)
• Advantage: No data modification required, preserving all original information
• Complement to SMOTETomek: Can be combined with resampling for layered imbalance handling

The area under the Precision-Recall curve is the preferred evaluation metric for highly imbalanced fraud datasets. Unlike ROC AUC, which can be misleadingly optimistic when negatives dominate, PR AUC focuses exclusively on the minority class's performance.

• Precision: Of all transactions flagged as fraud, how many were actually fraudulent?
• Recall: Of all actual fraud, how many did the model catch?
• Why It Matters: A model with 99.9% accuracy on a 0.1% fraud rate dataset is useless; PR AUC reveals true detection capability