A new imbalanced data oversampling method based on Bootstrap method and Wasserstein Generative Adversarial Network.

Area of Science:

• Bioinformatics
• Machine Learning
• Data Science

Background:

• Class imbalance in biological data poses challenges for traditional machine learning classifiers.
• Generative Adversarial Networks (GANs) show promise for imbalanced data but struggle with unknown minority class distributions.
• Standard GANs using random noise input can lead to training difficulties and lower quality generated data.

Purpose of the Study:

• To develop an improved oversampling algorithm for imbalanced biological datasets.
• To enhance the quality and realism of synthesized minority class samples.
• To improve the classification performance of machine learning models on imbalanced biological data.

Main Methods:

• Proposed a novel oversampling algorithm combining the Bootstrap method and Wasserstein GAN (BM-WGAN).
• Utilized Bootstrap method to estimate the minority class data distribution.
• Fed minority class data distribution into the generator for synthesizing new samples.

Main Results:

• The proposed BM-WGAN algorithm effectively synthesizes realistic minority class samples.
• BM-WGAN significantly improves classification performance compared to existing oversampling techniques.
• The generator learns valuable features from the minority class data.

Conclusions:

• BM-WGAN offers a robust solution for handling class imbalance in biological data classification.
• The method enhances the utility of GANs for imbalanced datasets by addressing input distribution challenges.
• BM-WGAN demonstrates superior performance, making it a valuable tool for bioinformatics research.