Self-Guided Deep Multiview Subspace Clustering via Consensus Affinity Regularization.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering (MVSC) utilizes complementary information from multiple data views.
• Existing MVSC methods struggle with high-dimensional, noisy raw features, leading to suboptimal clustering.
• There is a need for advanced MVSC techniques that handle feature noise and dimensionality effectively.

Purpose of the Study:

• To propose a novel Self-Guided Deep Multiview Subspace Clustering (SDMSC) model.
• To perform joint deep feature embedding and subspace analysis for robust clustering.
• To improve consensus data affinity relationship recovery by integrating multiple embedding spaces.

Main Methods:

• SDMSC integrates deep feature embedding with subspace analysis.
• It establishes a consensus data affinity relationship across all views and intermediate embedding spaces.
• Self-guided learning uses raw feature affinity as supervision to mitigate local minima during embedding.

Main Results:

• The proposed SDMSC model significantly outperforms state-of-the-art clustering methods.
• Experiments on seven datasets validate the effectiveness of the joint deep feature embedding and subspace analysis approach.
• The self-guided strategy enhances the reliability of the recovered data affinity relationship.

Conclusions:

• SDMSC offers a robust solution for multiview subspace clustering by addressing limitations of raw feature analysis.
• The self-guided deep embedding mechanism improves clustering accuracy and stability.
• The method demonstrates superior performance across diverse datasets, advancing the field of multiview learning.