Robust Least Squares Regression for Subspace Clustering: A Multi-View Clustering Perspective.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Subspace clustering (SC) methods assume data self-reconstruction and achieve success.
• SC methods often require parameter tuning, leading to varied affinity matrices.
• Existing SC methods do not leverage the complementary information across different parameter-tuned affinity matrices.

Purpose of the Study:

• To propose a novel method for fusing multiple affinity matrices generated by subspace clustering (SC) from a multi-view clustering (MVC) perspective.
• To enhance clustering performance by treating different affinity matrices as consistent and complementary views.
• To introduce a robust least squares regression from an MVC perspective (RLSR/MVCP).

Main Methods:

• Utilizing least squares regression (LSR) with varying parameters to generate multiple affinity matrices.
• Fusing these affinity matrices into a tensor, constrained by tensor nuclear norm (TNN) for noise reduction and information exploration.
• Solving the combined framework using the augmented Lagrange multiplier (ALM) method.

Main Results:

• The proposed RLSR/MVCP method demonstrates superior clustering performance compared to state-of-the-art SC methods.
• Experimental results on multiple datasets validate the effectiveness of the tensor fusion approach.
• The method successfully integrates information from different affinity matrices to improve robustness.

Conclusions:

• The proposed RLSR/MVCP framework effectively fuses multiple affinity matrices from SC using an MVC perspective.
• This approach enhances clustering accuracy and robustness by exploiting complementary information and reducing noise.
• The method represents a significant advancement in subspace clustering techniques.