A Robust Semi-Supervised Broad Learning System Guided by Ensemble-Based Self-Training.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Semi-supervised learning in Broad Learning Systems (BLS) aims to reduce label dependency.
• Existing self-training methods struggle with imbalanced data and concept drift.

Purpose of the Study:

• To propose a robust semi-supervised BLS using ensemble-based self-training (ESTSS-BLS).
• To address limitations of current methods in handling imbalanced data and concept drift.

Main Methods:

• Ensemble-based self-training determines pseudo-labels using multiple BLSs for improved reliability.
• Label purity metric ensures credibility of auxiliary training data.
• A data-driven dynamic nodes mechanism adjusts network structure to mitigate concept drift.

Main Results:

• ESTSS-BLS demonstrates superior performance over existing methods in accuracy, precision, recall, F1 score, and AUC.
• Achieved 87.84% accuracy on MNIST with only 0.1% labeled data.
• Matched fully supervised performance on NORB using just 2% labeled data.
• Showed stable performance on medical and biological datasets.

Conclusions:

• ESTSS-BLS offers a robust and adaptable solution for semi-supervised learning, particularly in challenging data scenarios.
• The method significantly reduces the need for labeled data while maintaining high accuracy.
• ESTSS-BLS proves effective across diverse datasets, including complex medical and biological data.