One-Step Multiview Subspace Segmentation via Joint Skinny Tensor Learning and Latent Clustering

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview subspace clustering (MSC) is valuable in NLP, face recognition, and time-series analysis.
• Existing MSC methods face challenges with high computational costs and multistage clustering.
• Current approaches often separate representation learning and clustering, limiting performance.

Purpose of the Study:

• To propose a novel MSC model, Joint Skinny Tensor Learning and Latent Clustering (JSTC).
• To address high computational cost and multistage clustering issues in MSC.
• To simultaneously learn high-order skinny tensor representations and latent clustering assignments.

Main Methods:

• Developed JSTC, a model integrating representation learning and clustering.
• Employed a joint optimization strategy to exploit multiview complementary information.
• Designed an alternating direction minimization algorithm for efficient optimization.

Main Results:

• JSTC demonstrated superior clustering performance across ten datasets.
• The method showed significant operational efficiency compared to 12 competitors.
• Statistical tests confirmed the method's effectiveness and efficiency.

Conclusions:

• JSTC offers an effective solution for large-scale MSC problems.
• The joint optimization strategy enhances the exploitation of multiview information.
• JSTC provides a computationally efficient and high-performing MSC approach.