A self-training algorithm based on the two-stage data editing method with mass-based dissimilarity.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Classical self-training algorithms in semi-supervised learning rely on limited labeled data and abundant unlabeled data.
• Existing methods often overlook data distribution when assessing sample similarity, focusing solely on geometric distance.
• Misclassified samples can significantly degrade the performance of self-training classifiers.

Purpose of the Study:

• To propose a novel self-training algorithm, STDEMB (self-training algorithm based on data editing with mass-based dissimilarity).
• To address limitations of existing methods by incorporating data distribution and improving sample similarity calculations.
• To enhance classifier accuracy by effectively handling misclassified samples during the training process.

Main Methods:

• Development of a mass matrix using mass-based dissimilarity to quantify sample relationships.
• Calculation of mass-based local density for each sample based on its k-nearest neighbors.
• Implementation of a two-stage data editing algorithm inspired by Density Peak Clustering (DPC) for sample refinement and selection.

Main Results:

• The STDEMB algorithm demonstrated effectiveness across 18 benchmark datasets.
• Experimental validation using accuracy and F-score metrics confirmed the algorithm's performance.
• The proposed method successfully addressed issues related to data distribution and misclassified samples.

Conclusions:

• The STDEMB algorithm offers a significant improvement over traditional self-training methods.
• Incorporating mass-based dissimilarity and data editing enhances the robustness and accuracy of semi-supervised learning.
• The study validates the efficacy of the proposed approach in improving classifier performance.