Semisupervised Multiple Choice Learning for Ensemble Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Ensemble learning enhances classification model performance.
• Semisupervised learning addresses challenges with partially labeled datasets.
• Existing methods may not fully exploit unlabeled data for domain-specific insights.

Purpose of the Study:

• To propose a novel semisupervised multiple choice learning (SemiMCL) approach.
• To jointly train network ensembles on partially labeled data.
• To improve labeled data assignment and utilize unlabeled data for domain-specific information.

Main Methods:

• Developed a SemiMCL framework for joint network ensemble training.
• Incorporated an auxiliary reconstruction task for domain-specific representation learning.
• Employed negative l1-norm regularization with conditional entropy minimization for implicit labeling of unlabeled data.

Main Results:

• The proposed SemiMCL model demonstrated effectiveness in semisupervised classification.
• Experiments on real-world datasets confirmed the model's superiority over existing methods.
• Improved labeled data assignment and better utilization of unlabeled data were observed.

Conclusions:

• The SemiMCL approach offers a robust solution for semisupervised classification with partially labeled data.
• The integration of an auxiliary task and novel regularization enhances representation learning.
• The model effectively leverages unlabeled data to boost predictive performance.