Structure regularized consensus dynamic anchor graph learning for incomplete multi-view clustering.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Incomplete multi-view clustering (IMVC) algorithms are crucial for data analysis but often overlook feature importance and structural information.
• Existing dynamic anchor graph-based IMVC methods have limitations in utilizing original feature space structures and individual feature weights.

Purpose of the Study:

• To propose a novel IMVC algorithm (SRCDAGL-IMC) that addresses limitations of existing methods.
• To leverage structural information from all views and incorporate sample-specific feature weights.
• To simultaneously recover missing features and learn a consensus anchor graph.

Main Methods:

• Developed SRCDAGL-IMC algorithm incorporating structural information as regularization terms.
• Introduced sample coefficients to weigh individual feature importance within each view.
• Employed an effective alternating optimization strategy for model training.
• Simultaneously performed consensus anchor graph learning and missing feature recovery.

Main Results:

• The proposed SRCDAGL-IMC algorithm demonstrated superior performance compared to state-of-the-art matrix factorization-based IMVC methods.
• Significant improvements were observed in terms of accuracy, normalized mutual information, and purity across six public datasets.
• The integration of structural information and feature weights proved effective in enhancing clustering quality.

Conclusions:

• SRCDAGL-IMC effectively addresses key limitations in existing IMVC algorithms.
• The proposed method offers a robust framework for incomplete multi-view clustering by integrating structural information and feature weighting.
• The findings suggest a promising direction for future research in multi-view learning and data imputation.