Multi-Label Classification via Adaptive Resonance Theory-Based Clustering.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multi-label classification assigns multiple labels to data instances.
• Continual learning enables models to adapt to new data without forgetting previous knowledge.
• Existing methods often struggle with dynamic label spaces and continuous adaptation.

Purpose of the Study:

• To propose a novel multi-label classification algorithm that supports continual learning.
• To integrate Adaptive Resonance Theory (ART) clustering with Bayesian probability for robust classification.
• To develop a method capable of handling an increasing number of labels dynamically.

Main Methods:

• Utilizing an ART-based clustering algorithm to generate adaptive prototype nodes for classification.
• Employing a Bayesian approach for label probability computation, counting label occurrences per class.
• Designing the system to independently update label probabilities, accommodating new labels.

Main Results:

• The proposed algorithm demonstrates competitive classification performance against established multi-label algorithms.
• Experimental validation on synthetic and real-world datasets confirms the algorithm's effectiveness.
• The system successfully implements continual learning, adapting to evolving data distributions.

Conclusions:

• The developed algorithm offers a viable solution for multi-label classification in dynamic environments.
• The combination of ART clustering and Bayesian probability computation enhances adaptability and performance.
• This approach provides a robust framework for machine learning systems requiring continual learning capabilities.