Radial-Based Oversampling for Multiclass Imbalanced Data Classification.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Learning from imbalanced data is a key machine learning challenge.
• Existing methods primarily focus on binary classification, leaving multiclass problems underexplored.
• Multiclass imbalanced datasets present unique challenges due to complex class relationships.

Purpose of the Study:

• To propose a novel data-sampling algorithm for multiclass imbalanced learning.
• To address the limitations of existing oversampling techniques in multiclass scenarios.
• To enhance the understanding and handling of relationships among multiple skewed classes.

Main Methods:

• Introduction of multiclass radial-based oversampling (MC-RBO).
• Utilizes potential functions for generating synthetic data instances.
• Considers information from all classes, unlike traditional methods focusing only on minority classes.
• Guides instance generation by identifying areas of low mutual class distribution.

Main Results:

• MC-RBO demonstrates significant improvements in learning from multiclass imbalanced data.
• The algorithm effectively handles difficult data distributions.
• Experimental studies and statistical analysis validate the algorithm's performance.

Conclusions:

• The proposed MC-RBO algorithm offers a superior approach to multiclass imbalanced learning.
• Considering all class information and employing smart oversampling are crucial for effective imbalanced data handling.
• MC-RBO alleviates shortcomings of existing methods, paving the way for better multiclass classification models.