An Incremental-Self-Training-Guided Semi-Supervised Broad Learning System.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Broad Learning System (BLS) is widely used but primarily supervised.
• Existing semi-supervised BLS methods have limitations with mixed labeled/unlabeled data.
• Need for improved BLS performance in practical applications with heterogeneous data.

Purpose of the Study:

• Propose an Incremental-Self-Training-guided Semi-Supervised BLS (ISTSS-BLS).
• Address limitations of traditional self-training and existing semi-supervised BLS.
• Enhance model performance and efficiency for mixed data scenarios.

Main Methods:

• Incremental Self-Training (IST) for acquiring unlabeled data.
• Double-restricted mechanism to prevent incorrect pseudo-labeling.
• Dynamic neuron-incremental mechanism for effective network structure updates.
• Iterative learning using a small labeled dataset and recursive self-updates.

Main Results:

• ISTSS-BLS demonstrates outstanding performance across 11 datasets.
• Incremental self-training saves up to 52.02% learning time compared to traditional methods.
• Proposed Accuracy-Time ratio (A/T) metric for comprehensive evaluation.
• ISTSS-BLS shows significant advantages over state-of-the-art alternatives.

Conclusions:

• ISTSS-BLS effectively handles mixed labeled and unlabeled data.
• The proposed methods ensure parsimonious model updates and prevent erroneous pseudo-labeling.
• ISTSS-BLS offers superior performance and efficiency in machine learning tasks.