SMOTEENN

SMOTEENN operates in a sequential, two-phase process:

• Stage 1 (SMOTE): Over-samples the minority class by creating synthetic examples through interpolation between a randomly selected minority instance and its k-nearest minority neighbors.
• Stage 2 (ENN): Applies Edited Nearest Neighbors to the entire augmented dataset. ENN removes any example whose class label differs from the majority class of its k-nearest neighbors (typically k=3). This pipeline ensures the final dataset has both increased minority representation and cleaner class boundaries.

The primary advantage of SMOTEENN over standalone SMOTE is its ability to remove overlapping and noisy instances introduced during synthetic generation.

• SMOTE can inadvertently create synthetic minority examples deep inside majority class clusters, introducing noise.
• The ENN cleaning step identifies and deletes these ambiguous examples, along with majority class instances that intrude on minority space.
• The result is a dataset with wider, more well-defined class separation margins, which is critical for algorithms like k-NN and SVM that are sensitive to boundary noise.

Unlike Tomek Links or One-Sided Selection, which primarily remove majority class examples, ENN cleans both classes:

• Majority class removal: Majority instances surrounded by minority neighbors are deleted, clarifying the minority boundary.
• Minority class removal: Synthetic or original minority instances that fall within dense majority regions are also removed, preventing the model from learning misleading patterns. This dual cleaning is particularly valuable in fraud detection, where legitimate transactions can exhibit anomalous patterns and fraudulent transactions can mimic normal behavior.

SMOTEENN is computationally more expensive than either SMOTE or ENN alone:

• SMOTE complexity: O(n_minority * k * d), where d is the feature dimensionality.
• ENN complexity: O(N * k * d), where N is the total number of examples in the augmented dataset.
• The ENN step must compute nearest neighbors for every instance in the expanded dataset, which can be substantial after SMOTE over-sampling. For large-scale financial transaction datasets with millions of records, consider using approximate nearest neighbor algorithms or applying SMOTEENN on a representative stratified sample.

SMOTEENN inherits hyperparameters from both constituent algorithms:

• SMOTE k-neighbors: Controls the number of nearest minority neighbors used for interpolation. Typical values range from 3 to 5.
• ENN k-neighbors: Determines how many neighbors vote on whether an instance should be removed. The default is 3; increasing this value makes the cleaning more aggressive.
• Sampling strategy: Dictates the desired minority-to-majority ratio after SMOTE (e.g., 0.5 for 1:2, 1.0 for perfect balance). Cross-validation should be performed on the full pipeline, not on individual components, as the interaction effects between SMOTE and ENN parameters significantly impact final model performance.

SMOTEENN is particularly effective in financial fraud anomaly detection pipelines:

• Preprocessing step: Applied before training classifiers like XGBoost, LightGBM, or Random Forest on imbalanced transaction data.
• Imbalanced-learn implementation: Available via SMOTEENN in the imblearn.combine module, with a scikit-learn-compatible fit_resample API.
• Pipeline integration: Can be embedded directly into scikit-learn Pipeline objects using imblearn.pipeline.Pipeline to prevent data leakage during cross-validation.
• Typical fraud ratios: Effective when fraud-to-legitimate ratios range from 1:100 to 1:1000, where standalone SMOTE may create excessive noise without the ENN cleaning step.

A combined resampling technique that first applies SMOTE to generate synthetic minority samples, then uses Tomek Links to remove overlapping majority class instances. Unlike SMOTEENN, which uses Edited Nearest Neighbors for cleaning, SMOTETomek identifies pairs of minimally distanced nearest neighbors of opposite classes and removes the majority class instance. This creates a cleaner, more defined class separation by eliminating borderline ambiguity rather than misclassified noise.

The cleaning component of SMOTEENN. ENN removes any example whose predicted class differs from its actual class based on its k-nearest neighbors (typically k=3). In SMOTEENN, ENN is applied after SMOTE over-sampling to eliminate:

• Noisy majority examples that intrude into minority space
• Ambiguous synthetic samples that may have been generated in overlapping regions
• Misclassified instances from both classes This dual cleaning creates a dataset with more homogeneous class clusters.

A more aggressive under-sampling algorithm that combines Edited Nearest Neighbors with additional minority class cleaning. NCR first applies ENN to remove majority class examples, then removes minority class examples that are misclassified by their three nearest neighbors. This two-pass approach cleans both classes simultaneously, making it more thorough than SMOTEENN's single ENN pass. NCR is particularly effective when the minority class itself contains outliers or mislabeled instances that could mislead the classifier.

A combined under-sampling strategy that first removes redundant majority class examples using Tomek Links, then applies Condensed Nearest Neighbors (CNN) to eliminate instances far from the decision boundary. Unlike SMOTEENN which over-samples first, OSS is purely an under-sampling approach that retains only the majority class examples essential for defining the decision boundary. This makes OSS more conservative in dataset size reduction but avoids introducing synthetic data.

An ensemble method that integrates SMOTE into the AdaBoost boosting procedure. At each boosting iteration, SMOTE generates new synthetic minority class examples before training the next weak learner. This increases the emphasis on the rare class progressively as boosting continues. Unlike SMOTEENN which produces a single cleaned dataset for one model, SMOTEBoost creates multiple balanced training sets across iterations, combining the benefits of synthetic sampling with ensemble diversity for improved generalization.

A variant of SMOTE that only over-samples minority class examples near the decision boundary, where misclassification is most likely. It classifies minority examples into three groups—safe, borderline, and noise—based on their neighbor composition, then generates synthetic samples only from borderline instances. SMOTEENN can be combined with Borderline-SMOTE as the over-sampling component to focus synthetic generation on the most critical regions before ENN cleaning, creating an even more targeted hybrid approach.