Enhanced manifold regularization for semi-supervised classification.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Manifold regularization (MR) is a popular semi-supervised learning technique that leverages data's local manifold structure.
• Existing MR methods primarily focus on data smoothness and overlook discriminative information, potentially limiting performance.
• Exploiting both manifold structure and discriminative properties is crucial for effective semi-supervised classification.

Purpose of the Study:

• To propose an enhanced manifold regularization (MR) framework for semi-supervised classification.
• To explicitly incorporate local discriminative information from both labeled and unlabeled data.
• To improve classification performance by addressing the limitations of traditional MR approaches.

Main Methods:

• A semi-supervised clustering method (semi-supervised fuzzy c-means) is used to uncover the intrinsic data structure.
• A local discrimination graph is constructed to model discriminative information, enforcing separation between potentially different clusters.
• The discrimination graph is integrated into the MR framework, combined with Laplacian regularized Kernel minimum squared error for classification.

Main Results:

• The proposed enhanced MR framework effectively utilizes local discriminative information alongside manifold structure.
• Experimental results on benchmark datasets and face recognition tasks demonstrate significant improvements in classification accuracy.
• The method successfully enforces separation between data points from different clusters, even if they appear similar on the manifold.

Conclusions:

• The enhanced MR framework provides a more comprehensive approach to semi-supervised classification by integrating discriminative learning.
• Explicitly modeling and utilizing local discriminative information leads to superior performance compared to standard MR.
• The proposed method offers a robust and effective solution for semi-supervised classification problems, particularly in domains like face recognition.