LVQ-SMOTE - Learning Vector Quantization based Synthetic Minority Over-sampling Technique for biomedical data

Area of Science:

• Biomedical Data Science
• Machine Learning
• Bioinformatics

Background:

• Imbalanced biomedical datasets pose challenges for traditional classification algorithms.
• Existing Synthetic Minority Over-sampling Technique (SMOTE) methods offer limited improvements over basic SMOTE.
• Vast empty feature spaces hinder accurate borderline estimation between classes.

Purpose of the Study:

• To develop a novel over-sampling method to enhance the classification of imbalanced biomedical data.
• To generate synthetic samples that occupy more feature space, improving upon existing SMOTE algorithms.
• To leverage learning vector quantization codebooks for more effective synthetic data generation.

Main Methods:

• A new over-sampling method was developed utilizing codebooks from learning vector quantization.
• The proposed method generates synthetic samples by referencing actual data points.
• Integration with existing SMOTE variants, such as MWMOTE, was explored.

Main Results:

• The novel over-sampling method demonstrated superior performance compared to basic SMOTE on four out of five standard classification algorithms across eight real-world imbalanced datasets.
• Performance gains were observed when the proposed method was combined with MWMOTE (Minority-over-sampling Weighted Majority Technique).
• Analysis of β-turn types prediction datasets revealed novel patterns previously unobserved.

Conclusions:

• The proposed over-sampling method effectively generates useful synthetic samples for imbalanced biomedical data classification.
• The method is compatible with standard classification algorithms and existing over-sampling techniques.
• This approach offers a promising solution for improving machine learning model accuracy in biomedical applications.