Non-Convex Sparse and Low-Rank Based Robust Subspace Segmentation for Data Mining.

Area of Science:

• Data Mining
• Machine Learning
• Computer Vision

Background:

• Parsimony, encompassing sparsity and low-rank properties, is crucial for social network data mining tasks like segmentation and recognition.
• Conventional methods using convex optimization (l₁-norm, nuclear norm) often yield suboptimal solutions for sparse and low-rank problems.

Purpose of the Study:

• To propose a novel robust subspace segmentation algorithm that overcomes the limitations of convex optimization.
• To enhance the capture of local geometrical structure and global data information for improved segmentation.

Main Methods:

• Integration of l_p-norm and Schatten p-norm constraints into a subspace segmentation model.
• Development of an efficient linearized alternating direction method for model optimization.

Main Results:

• The proposed algorithm generates an affinity graph that better represents data structure.
• Experimental results on public datasets demonstrate superior effectiveness and robustness compared to five existing algorithms.

Conclusions:

• The novel algorithm provides a more generative, discriminative, and robust approach to subspace segmentation.
• The integration of non-convex norms offers significant advantages over traditional convex methods in data mining applications.