Robust Nonnegative Matrix Factorization With Self-Initiated Multigraph Contrastive Fusion.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Graph regularized nonnegative matrix factorization (GNMF) is effective for dimensionality reduction but struggles with noisy or polluted datasets.
• Outliers in data, such as obscured faces, can lead to inaccurate graph representations and poor clustering results in standard GNMF models.

Purpose of the Study:

• To propose a novel robust nonnegative matrix factorization method (RNMF-SMGF) for improved subspace learning and clustering on polluted data.
• To enhance the accuracy of graph regularization by fusing multiple self-initiated graph structures.

Main Methods:

• Introduced Robust Nonnegative Matrix Factorization with Self-Initiated Multi-Graph Contrastive Fusion (RNMF-SMGF).
• Developed a self-initiated approach to learn diverse graph structures from different data perspectives without altering original data.
• Integrated entropy regularization and L2,1/2-norm constraints for robust learning and effective cluster formation.

Main Results:

• RNMF-SMGF demonstrated superior performance in robust clustering tasks compared to existing methods.
• The proposed method effectively handles polluted data and mitigates the impact of outliers.
• Experiments on benchmark datasets validated the model's effectiveness in subspace learning and clustering.

Conclusions:

• RNMF-SMGF offers a robust solution for clustering challenging datasets with outliers.
• The multi-graph contrastive fusion strategy significantly improves representation learning accuracy.
• The method provides a reliable approach for unsupervised subspace learning in real-world applications.