A Generative Neighborhood-Based Deep Autoencoder for Robust Imbalanced Classification.

Area of Science:

• Machine Learning
• Deep Learning
• Data Science

Background:

• Deep learning models require large, balanced datasets for optimal performance.
• Real-world applications often suffer from limited, imbalanced data, leading to poor classification.
• Existing imbalanced learning methods are often application-specific or require expert knowledge.

Purpose of the Study:

• To address the limitations of current imbalanced data classification methods.
• To introduce a simple yet effective generative model applicable to both image and time-series data.
• To improve prediction stability and performance in imbalanced learning scenarios.

Main Methods:

• Developed GENDA, a generative neighborhood-based deep autoencoder.
• GENDA learns latent representations using the neighboring embedding space of samples.
• The model is designed for broad applicability across different data types.

Main Results:

• GENDA demonstrates efficacy on various real-world imbalanced datasets.
• The method shows successful application to both image and time-series data.
• Achieved improved results even under severe data imbalance ratios.

Conclusions:

• GENDA offers a versatile and effective solution for imbalanced data classification.
• The proposed generative model overcomes limitations of existing approaches.
• GENDA is a competitive and accessible method for diverse applications.