CoTrade: Confident Co-Training With Data Editing.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Co-training is a semi-supervised learning method using two classifiers on different data views.
• Traditional co-training can be unstable due to erroneous label propagation between classifiers with mediocre accuracy.
• Reliable label communication is crucial for effective co-training performance.

Purpose of the Study:

• To address the instability issue in co-training algorithms.
• To propose a novel co-training algorithm, COTRADE, for reliable label communication.
• To enhance the generalization performance of semi-supervised learning models.

Main Methods:

• COTRADE employs a two-step label communication process in each round.
• Explicitly estimates classifier prediction confidence using data editing techniques.
• Passes high-confidence predicted labels between classifiers with imposed constraints to prevent noise.

Main Results:

• COTRADE demonstrates effective exploitation of unlabeled data.
• Experiments show improved generalization performance across real-world datasets.
• The algorithm successfully reduces undesirable classification noise during co-training.

Conclusions:

• COTRADE offers a more stable and reliable approach to co-training.
• The method effectively mitigates the issue of erroneous label propagation.
• COTRADE enhances the utility of unlabeled data in semi-supervised learning scenarios.