Constrained Clustering With Dissimilarity Propagation-Guided Graph-Laplacian PCA

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Constrained clustering leverages pairwise constraints to guide data partitioning.
• Existing methods often struggle to effectively integrate both local and global data structures using limited supervisory information.

Purpose of the Study:

• To propose a novel constrained clustering model, DP-GLPCA, that effectively utilizes pairwise constraints.
• To enhance clustering by capturing both local and global data structures.

Main Methods:

• Developed a convex semisupervised low-dimensional embedding model by integrating a dissimilarity regularizer into Graph-Laplacian Principal Component Analysis (GLPCA).
• Introduced a novel dissimilarity propagation technique to refine the dissimilarity regularizer using 'cannot-link' constraints.
• Designed an efficient iterative algorithm based on the inexact augmented Lagrange multiplier for model optimization, with guaranteed global convergence.

Main Results:

• DP-GLPCA achieved significantly higher clustering accuracy compared to state-of-the-art methods across nine benchmark datasets.
• Experimental validation confirmed the effectiveness and advantages of the proposed dissimilarity propagation model.
• Demonstrated the first investigation into dissimilarity propagation for constrained clustering.

Conclusions:

• The proposed DP-GLPCA model offers superior performance in constrained clustering by effectively integrating local and global data structures.
• Dissimilarity propagation represents a novel and effective approach for enhancing constrained clustering methods.
• The method provides a robust framework for leveraging pairwise constraints in dimensionality reduction and clustering tasks.