Multiobjective Semisupervised Classifier Ensemble.

Area of Science:

• Machine Learning
• Data Mining
• Bioinformatics

Background:

• Classifying high-dimensional data with scarce labels presents significant challenges in machine learning.
• Existing methods often struggle with feature relevance and redundancy in such datasets.

Purpose of the Study:

• To develop an advanced semisupervised classification approach for high-dimensional data.
• To enhance classification performance by optimizing feature subspace selection and ensemble training.

Main Methods:

• Proposed the multiobjective semisupervised classifier ensemble (MOSSCE) framework.
• Introduced a multiobjective subspace selection process (MOSSP) with objectives for feature relevance, redundancy, and reconstruction error.
• Generated an auxiliary training set using sample confidence to boost classifier ensemble performance.

Main Results:

• MOSSCE demonstrated superior performance compared to state-of-the-art semisupervised classifiers.
• Experiments on 12 gene expression and 2 large image datasets validated the approach's effectiveness.
• Diversity analysis and statistical tests confirmed the robustness of the MOSSCE method.

Conclusions:

• MOSSCE effectively addresses the challenge of classifying high-dimensional data with limited labels.
• The proposed method offers a significant improvement in classification accuracy and robustness.
• This approach holds promise for applications in bioinformatics and image analysis.