Addressing imbalanced data classification with Cluster-Based Reduced Noise SMOTE.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced data presents a significant challenge in machine learning, negatively impacting classification algorithm performance.
• Existing oversampling methods often struggle with noise reduction and maintaining class separability.

Purpose of the Study:

• To introduce Cluster-Based Reduced Noise SMOTE (CRN-SMOTE), a novel data-level oversampling technique.
• To enhance classification performance on imbalanced datasets through effective noise reduction and minority class oversampling.

Main Methods:

• CRN-SMOTE integrates SMOTE (Synthetic Minority Over-sampling Technique) with a unique cluster-based noise reduction strategy.
• The noise reduction technique ensures that samples from each class form distinct clusters, a feature not achieved by conventional methods.
• Evaluation was performed on four imbalanced datasets (ILPD, QSAR, Blood, Maternal Health Risk) using key metrics like Cohen's kappa, MCC, F1-score, precision, and recall.

Main Results:

• CRN-SMOTE consistently outperformed state-of-the-art methods including RN-SMOTE, SMOTE-Tomek Link, and SMOTE-ENN across all tested datasets.
• Significant performance gains were observed on the QSAR and Maternal Health Risk datasets.
• CRN-SMOTE achieved 100% superiority over RN-SMOTE, with average improvements of 6.6% in Kappa, 4.01% in MCC, 1.87% in F1-score, 1.7% in precision, and 2.05% in recall when SMOTE's neighbors were set to 5.

Conclusions:

• CRN-SMOTE offers a superior approach to handling imbalanced data compared to existing methods.
• The proposed cluster-based noise reduction is key to CRN-SMOTE's effectiveness in improving classification accuracy.
• This method demonstrates strong potential for enhancing machine learning model performance in real-world imbalanced classification scenarios.