Semi-supervised learning for ordinal Kernel Discriminant Analysis.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Ordinal classification problems, where predicted labels have a natural order, often suffer from scarce labeled data.
• Obtaining ordinal labels typically requires human experts, making large labeled datasets difficult to acquire, especially for recommendation systems.

Purpose of the Study:

• To develop a new semi-supervised learning strategy for ordinal classification that leverages both labeled and unlabeled data.
• To devise a novel semi-supervised kernel learning method tailored for ordinal classification to optimize kernel parameters.

Main Methods:

• Extending the ordinal version of kernel discriminant learning to incorporate neighborhood information from unlabeled data.
• Computing distances in the feature space induced by the kernel function to enhance the utilization of unlabeled data.
• Combining a new semi-supervised kernel learning method with the developed classification strategy for kernel parameter optimization.

Main Results:

• The proposed ordinal classification strategy shows strong synergy with semi-supervised learning, effectively utilizing unlabeled data.
• Computing distances in the kernel-induced feature space provides a significant advantage for semi-supervised ordinal classification.
• Experimental comparisons across 30 datasets validate the effectiveness of the proposed methods against six other semi-supervised approaches.

Conclusions:

• The integration of ordinal discriminant analysis with semi-supervised learning is highly effective for classification tasks with ordered labels.
• Utilizing unlabeled data and computing distances in the kernel feature space are crucial for advancing semi-supervised ordinal classification.