Meta-cognitive online sequential extreme learning machine for imbalanced and concept-drifting data classification.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Class imbalance and concept drift are significant challenges in real-world data stream learning.
• Existing sequential learning methods often struggle with both issues simultaneously, especially in multi-class scenarios.
• Online sequential extreme learning machines (OS-ELM) offer a framework for rapid learning but require adaptation for complex data stream problems.

Purpose of the Study:

• To propose a novel meta-cognitive online sequential extreme learning machine (MOS-ELM) capable of handling class imbalance and concept drift.
• To develop a self-regulating learning mechanism within MOS-ELM that adaptively selects appropriate strategies.
• To establish MOS-ELM as the first sequential learning approach to address class imbalance in both binary and multi-class data streams concurrently with concept drift.

Main Methods:

• Introduction of meta-cognition for adaptive strategy selection in MOS-ELM.
• Development of a new adaptive window approach specifically designed for concept drift detection and learning.
• Formulation of a unified single output update equation to generalize OS-ELM applications.

Main Results:

• MOS-ELM demonstrated superior performance across various conditions compared to specialized methods.
• The proposed adaptive window approach effectively managed concept drift in data streams.
• The unified output equation simplified and enhanced the applicability of OS-ELM.

Conclusions:

• MOS-ELM provides an effective and unified solution for learning from imbalanced data streams with concept drift.
• The meta-cognitive approach enables robust self-regulation, leading to improved learning performance.
• This work advances sequential learning techniques for complex, dynamic data environments.