Semi-supervised learning for tree-structured ensembles of RBF networks with Co-Training

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Supervised learning demands extensive labeled data, which is costly and time-consuming to acquire.
• Co-Training is a semi-supervised method that leverages unlabeled data to reduce labeling needs and improve classification.
• Multi-class problems with numerous categories pose significant challenges for traditional methods.

Purpose of the Study:

• To develop and evaluate new learning architectures that integrate tree-structured approaches with Co-Training.
• To address the limitations of single-view tree-structured methods in large-scale multi-class classification tasks.
• To improve recognition accuracy in scenarios with limited labeled data by exploiting unlabeled data and multiple feature views.

Main Methods:

• Proposed two novel learning architectures combining tree-structured decomposition and Co-Training.
• Exploited redundantly sufficient and conditionally independent feature sets (views) for classification.
• Applied output space decomposition to break down complex multi-class problems into binary sub-problems.

Main Results:

• Demonstrated the effectiveness of the proposed architectures for classification tasks with many classes and scarce labeled data.
• Showcased improved recognition accuracy compared to single-view tree-structured approaches when combined with Co-Training.
• Validated the ability of the combined methods to effectively utilize independent views and unlabeled data.

Conclusions:

• The proposed tree-structured Co-Training architectures offer a powerful solution for complex multi-class classification with limited labeled data.
• Combining Co-Training with tree-structured methods enhances the exploitation of unlabeled data and feature views.
• These architectures significantly improve classification performance in challenging real-world pattern recognition scenarios.