Semi-Supervised k-Star (SSS): A Machine Learning Method with a Novel Holo-Training Approach.

Area of Science:

• Machine Learning
• Information Theory
• Data Mining

Background:

• The k-Star algorithm, an entropy-based classification method, excels in performance and generalization.
• Standard k-Star is a supervised learning method, relying solely on labeled data.
• Limitations exist in supervised learning when labeled data is scarce.

Purpose of the Study:

• To introduce an improved Semi-Supervised k-Star (SSS) algorithm that incorporates unlabeled data.
• To propose a novel holo-training approach for semi-supervised learning.
• To enhance classification accuracy and model robustness in machine learning tasks.

Main Methods:

• Developed the Semi-Supervised k-Star (SSS) algorithm, integrating labeled and unlabeled data.
• Introduced holo-training, a new semi-supervised learning strategy.
• Employed entropy measures and combined multiple classifiers for robust modeling.

Main Results:

• The holo-training approach outperformed self-training on 13 of 18 datasets.
• The SSS method achieved an average accuracy of 95.25%, surpassing state-of-the-art methods (90.01%).
• Statistical validation using Binomial Sign and Friedman tests confirmed result significance.

Conclusions:

• The proposed SSS method and holo-training approach offer significant advancements in semi-supervised learning.
• Integrating unlabeled data effectively boosts classification performance.
• The SSS algorithm provides a powerful and robust solution for classification tasks.