Maximum Correntropy Criterion-Based Hierarchical One-Class Classification.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• One-class classification algorithms are vital for anomaly/outlier detection.
• Existing methods like OC-SVM and OC-ELM often use minimum mean-square-error (MSE), which struggles with large outliers.
• Deep learning approaches such as Deep SVDD and ML-OCELM have been developed but share similar limitations.

Purpose of the Study:

• To propose a robust Maximum Correntropy Criterion (MCC)-based One-Class Extreme Learning Machine (MC-OCELM).
• To extend MC-OCELM into a hierarchical network (HC-OCELM) for complex and large datasets.
• To improve outlier detection performance compared to existing methods.

Main Methods:

• Developed MC-OCELM utilizing the robust MCC criterion.
• Extended MC-OCELM to HC-OCELM for enhanced data characterization.
• Employed gradient derivation and fixed-point iteration for output weight optimization.

Main Results:

• MC-OCELM and HC-OCELM demonstrated effectiveness in anomaly detection.
• The proposed methods showed superior performance in handling large outliers.
• Experimental validation on benchmark datasets confirmed the algorithms' capabilities.

Conclusions:

• The proposed MC-OCELM and HC-OCELM offer robust solutions for anomaly detection.
• These algorithms overcome the limitations of MSE-based methods in dealing with significant outliers.
• The hierarchical extension enhances capability for complex, large-scale data analysis.