Spectral Clustering by Joint Spectral Embedding and Spectral Rotation

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Spectral clustering is vital for pattern recognition and image segmentation.
• Classical methods use separate stages for continuous and binary matrix computation, potentially losing information.
• Existing joint models struggle with approximating non-orthonormal matrices, leading to approximation errors.

Purpose of the Study:

• To develop a joint spectral clustering model that simultaneously computes optimal real and binary matrices.
• To address the approximation errors in current joint models by introducing a scaled orthonormal cluster indicator matrix.
• To improve the performance and accuracy of spectral clustering algorithms.

Main Methods:

• Proposing a joint model for simultaneous computation of optimal real and binary matrices.
• Replacing the non-orthonormal cluster indicator matrix with a scaled orthonormal cluster indicator matrix.
• Minimizing the difference between two orthonormal matrices for easier optimization.

Main Results:

• The proposed method, JSESR, demonstrates superior performance compared to existing techniques.
• Experimental results on benchmark datasets validate the effectiveness of the new approach.
• The joint optimization strategy leads to more accurate clustering outcomes.

Conclusions:

• The JSESR method effectively overcomes the limitations of previous joint spectral clustering models.
• Utilizing a scaled orthonormal cluster indicator matrix significantly improves clustering accuracy.
• The proposed approach offers a more robust and efficient solution for pattern recognition and image segmentation tasks.