Online sequential class-specific extreme learning machine for binary imbalanced learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Class imbalance is a prevalent issue in real-world machine learning applications, where minority classes have significantly fewer samples.
• Online sequential learning methods are effective for handling large-scale imbalanced classification problems.
• Existing methods like weighted online sequential extreme learning machine (WOS-ELM) and class-specific extreme learning machine (CS-ELM) have shown promise but can be improved.

Purpose of the Study:

• To propose a novel online sequential class-specific extreme learning machine (OSCSELM) algorithm.
• To enhance the effectiveness of handling class imbalance in both small and large datasets using online learning.
• To reduce computational complexity compared to existing WOS-ELM methods.

Main Methods:

• The proposed OSCSELM is a variant of CS-ELM that incorporates online sequential learning capabilities.
• OSCSELM supports both chunk-by-chunk and one-by-one online learning modes.
• The method utilizes class-specific regularization within an online learning framework.

Main Results:

• The OSCSELM method demonstrates superior performance in handling imbalanced learning compared to other benchmarked methods.
• Experimental results on real-world imbalanced datasets validate the effectiveness of the proposed approach.
• OSCSELM exhibits lower computational complexity than WOS-ELM.

Conclusions:

• The novel OSCSELM effectively addresses the class imbalance problem in online sequential learning scenarios.
• OSCSELM provides a computationally efficient and high-performing solution for imbalanced classification across various dataset sizes.
• This work advances online learning techniques for tackling challenging real-world data distributions.