Robust Bi-Stochastic Graph Regularized Matrix Factorization for Data Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Data clustering is crucial for data analysis, with Non-negative Matrix Factorization (NMF) being a powerful technique.
• Standard NMF and existing robust variants are sensitive to noise and outliers.
• Graph-based NMF methods often rely on fixed initial graphs and separate graph construction and factorization steps, leading to suboptimal performance.

Purpose of the Study:

• To propose a novel Robust Bi-Stochastic Graph Regularized Matrix Factorization (RBSMF) framework.
• To address the limitations of existing NMF methods, including sensitivity to noise and suboptimal graph learning.
• To develop a more effective and robust data clustering approach.

Main Methods:

• Developed a general, robust loss function superior to L2 and L1 norms.
• Introduced an adaptive similarity graph learning mechanism.
• Integrated graph updating and matrix factorization into a simultaneous process.

Main Results:

• The proposed RBSMF framework demonstrates enhanced robustness against noise and outliers.
• Simultaneous learning of the graph and factorization leads to more appropriate graph structures for clustering.
• Extensive experiments confirm that RBSMF outperforms state-of-the-art clustering methods.

Conclusions:

• RBSMF offers a significant advancement in robust data clustering.
• The simultaneous optimization of graph structure and factorization is key to improved performance.
• This framework provides a more effective solution for partitioning data into meaningful groups.