ADASYN

ADASYN is readily available in the imbalanced-learn Python library, compatible with scikit-learn pipelines. Key hyperparameters include sampling_strategy (target class ratio), n_neighbors (default 5 for density estimation), and random_state for reproducibility. The implementation uses the same interpolation mechanism as SMOTE but applies the adaptive density weighting before generation, making it a drop-in replacement for standard oversampling in imbalanced classification workflows.

The foundational over-sampling algorithm that ADASYN builds upon. Instead of duplicating minority examples, SMOTE creates synthetic samples by interpolating between a selected point and its randomly chosen k-nearest neighbors in feature space. This expands the minority class decision region, forcing the classifier to generalize better than with simple random oversampling. While effective, standard SMOTE treats all minority examples equally and can amplify noise if generated between outliers.

A targeted variant that directly addresses ADASYN's core philosophy of focusing on 'hard' examples. Borderline-SMOTE only over-samples minority instances near the decision boundary, ignoring safe interior points. It classifies minority examples into danger, safe, and noise categories based on the class distribution of their neighbors. By synthesizing data only in the region of overlap, it sharpens the class boundary and reduces the computational waste of generating trivial synthetic samples.

A powerful hybrid pipeline that combines over-sampling with data cleaning. SMOTETomek first applies SMOTE to generate synthetic minority samples, then uses Tomek Links to remove overlapping majority class instances. A Tomek Link is a pair of minimally distanced neighbors of opposite classes; removing the majority instance clarifies the boundary. This two-step process creates well-defined, non-overlapping class clusters that significantly improve classifier performance on dense, messy financial data.

An algorithmic-level alternative to data resampling. Instead of balancing the dataset, cost-sensitive learning penalizes the model more heavily for misclassifying the minority class. This is implemented via a cost matrix in algorithms like SVM or through hyperparameters like scale_pos_weight in XGBoost. The model optimizes for minimum misclassification cost rather than error rate, effectively biasing the decision boundary toward the rare class without generating any synthetic data.

A loss function designed to address extreme class imbalance during training without explicit resampling. Focal Loss adds a modulating factor to the standard cross-entropy loss that down-weights the contribution of easily classified examples. The formula FL = -(1 - p_t)^γ * log(p_t) forces the model to focus training on a sparse set of hard, misclassified samples. With a tunable focusing parameter (γ), it naturally replicates the adaptive emphasis that ADASYN achieves through data generation.

An ensemble method that integrates SMOTE directly into the AdaBoost boosting procedure. At each boosting iteration, SMOTEBoost generates new synthetic minority examples before training the next weak learner. This ensures that each subsequent classifier sees a balanced dataset and places increasing emphasis on the rare class. Unlike ADASYN's single-pass generation, this iterative approach progressively refines the decision boundary, often yielding higher recall on complex fraud patterns.