Tri-Training Algorithm for Adaptive Nearest Neighbor Density Editing and Cross Entropy Evaluation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Tri-training improves classifier generalization by pseudo-labeling unlabeled data.
• However, mislabeled data introduces noise, degrading classifier efficiency and accuracy.
• Existing methods struggle with noise reduction and accurate prediction in tri-training.

Purpose of the Study:

• To introduce the Tri-training algorithm for Adaptive Nearest Neighbor Density Editing and Cross-Entropy evaluation (TTADEC).
• To mitigate training noise and enhance prediction accuracy in semi-supervised learning.
• To address limitations of explicit decision mechanisms in tri-training.

Main Methods:

• TTADEC employs nearest neighbor editing for high-confidence sample labeling.
• Relative nearest neighbors define local density for pre-training sample screening.
• Adaptive techniques dynamically expand the training set.
• Cross-entropy evaluates base classifiers, assigning weights for a robust decision function.

Main Results:

• TTADEC effectively reduces training noise during classifier iteration.
• The algorithm demonstrates improved classification performance on the UCI dataset.
• Experimental results show TTADEC outperforms standard tri-training and its variants.

Conclusions:

• TTADEC offers a robust solution for semi-supervised classification with insufficient training data.
• The adaptive nearest neighbor density editing and cross-entropy evaluation significantly improve model accuracy.
• TTADEC effectively handles training noise, enhancing overall classifier performance.