Manifold regularized semi-supervised Gaussian mixture model.

Area of Science:

• Data Mining
• Pattern Recognition
• Machine Learning

Background:

• Gaussian Mixture Models (GMMs) are widely used for data clustering.
• The Expectation-Maximization (EM) algorithm is standard for GMM parameter estimation.
• Locally Consistent GMM (LCGMM) improves clustering by leveraging data's local manifold structure.

Purpose of the Study:

• To introduce a Semi-Supervised LCGMM (Semi-LCGMM) that integrates prior knowledge via partial class labels.
• To enhance clustering performance by incorporating supervised information into the LCGMM framework.
• To enable each class to be modeled by multiple Gaussian components, unlike unsupervised methods.

Main Methods:

• Developed Semi-LCGMM by incorporating prior class label knowledge into the LCGMM maximum likelihood function.
• Utilized the Expectation-Maximization (EM) algorithm for parameter estimation in the Semi-LCGMM.
• Leveraged the p nearest neighbor graph to exploit the local manifold structure of the data.

Main Results:

• Demonstrated promising clustering results across various applications.
• Successfully applied Semi-LCGMM to diverse datasets including medical data, image data, and segmentation tasks.
• Showcased the effectiveness of incorporating partial supervision for improved clustering outcomes.

Conclusions:

• Semi-LCGMM offers a robust approach to semi-supervised clustering by effectively utilizing partial class labels.
• The method enhances traditional LCGMM by allowing more flexible class modeling (multiple components per class).
• Semi-LCGMM shows significant potential for real-world applications requiring accurate data clustering with limited supervision.