In comparative evaluations on 44 KEEL datasets, the newly proposed Feature-Weighted Borderline-SMOTE (FWB-SMOTE) algorithm outperforms 11 benchmark oversampling algorithms on Decision Tree (DT), Support Vector Machine (SVM), and k-Nearest Neighbor (kNN) classifiers. The study, conducted by Mingyuan Liu and Qicheng Liu and published in Cluster Computing, evaluated classification performance using Area Under the Curve (AUC) and G-mean indicators to verify the effectiveness of the approach in improving synthetic minority class sample generation.
The researchers developed FWB-SMOTE to resolve key structural limitations in the classic Synthetic Minority Oversampling Technique (SMOTE) used for handling imbalanced datasets. While SMOTE remains a standard approach, it introduces noisy samples, blurs class boundaries, and equalizes feature weights during the generation of minority class samples. These deficiencies motivated the design of an algorithm capable of enhancing oversampling quality through systematic noise filtering and feature importance weighting.
In the initial procedural phase of the proposed algorithm, the framework calculates the Euclidean distance matrix between minority class samples and the entire dataset. FWB-SMOTE then filters out noisy samples by combining this distance calculation with an assessment of the proportion of neighboring sample categories. This initial filtering step ensures that noisy instances are eliminated prior to subsequent feature processing and synthetic generation.
The algorithm then directly quantifies feature importance by utilizing the feature splitting gain of XGBoost. It normalizes these calculated importance weights and maps them to the individual feature dimensions. This calibrated weighting mechanism ensures that key features dominate during the generation of synthetic minority samples while actively suppressing interference caused by redundant features.
In the final stage, the k-nearest neighbor algorithm determines minority class boundary samples within the weighted feature space, and the selected boundary samples are oversampled by integrating the assigned feature weights. Tested across 44 KEEL benchmark datasets, the FWB-SMOTE method achieved superior AUC and G-mean results across DT, SVM, and kNN classifiers compared to all 11 benchmark algorithms, confirming its effectiveness for class imbalance problems.
Subject of Research: Feature-weighted borderline oversampling for class imbalance problems
Article Title: FWB-SMOTE: feature-weighted borderline synthetic minority oversampling technique for class imbalance problems
Article References: Liu, M., & Liu, Q. (2027). FWB-SMOTE: feature-weighted borderline synthetic minority oversampling technique for class imbalance problems. Cluster Computing, 30(1), Article 25. https://doi.org/10.1007/s10586-026-06659-9
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06659-9
Keywords: FWB-SMOTE, SMOTE, class imbalance, oversampling, feature weighting, borderline-SMOTE, XGBoost, feature splitting gain, Euclidean distance matrix, noise filtering, k-nearest neighbor, decision tree, support vector machine, KEEL datasets, AUC indicator, G-mean indicator, minority class samples, boundary samples
Cite Scienmag News
Denise Maddox. (November 16, 2026). FWB-SMOTE Algorithm Outperforms Eleven Benchmarks on Imbalanced Datasets Across Three Classifiers. Scienmag. https://scienmag.com/fwb-smote-algorithm-outperforms-eleven-benchmarks-on-imbalanced-datasets-across-three-classifiers/
Denise Maddox. "FWB-SMOTE Algorithm Outperforms Eleven Benchmarks on Imbalanced Datasets Across Three Classifiers." Scienmag, 16 November 2026, https://scienmag.com/fwb-smote-algorithm-outperforms-eleven-benchmarks-on-imbalanced-datasets-across-three-classifiers/. Accessed 16 November 2026.
Denise Maddox. "FWB-SMOTE Algorithm Outperforms Eleven Benchmarks on Imbalanced Datasets Across Three Classifiers." Scienmag. November 16, 2026. https://scienmag.com/fwb-smote-algorithm-outperforms-eleven-benchmarks-on-imbalanced-datasets-across-three-classifiers/








