Deep Spectral Clustering With Constrained Laplacian Rank.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Spectral clustering (SC) is a widely used clustering technique.
• Existing SC methods face challenges with scalability and generalization for large datasets.
• The out-of-sample extension problem limits SC's applicability.

Purpose of the Study:

• To propose an efficient deep clustering architecture based on spectral clustering.
• To address the scalability and generalization limitations of current deep spectral clustering (DSC) methods.
• To develop a novel method, Deep SC with Constrained Laplacian Rank (DSCCLR), for improved clustering performance.

Main Methods:

• Developed DSCCLR, a deep clustering architecture incorporating constrained Laplacian rank.
• Created a self-adaptive affinity matrix with a clustering-friendly structure by constraining the Laplacian rank.
• Introduced a fully connected network with an orthogonality constraint for efficient representation learning.

Main Results:

• DSCCLR overcomes the limited generalization ability and scalability issues of existing DSC methods.
• The method effectively explores intrinsic sample relationships within the affinity matrix, preserving the data's latent manifold.
• DSCCLR alleviates the computational complexity associated with eigendecomposition through an efficient fully connected network.
• Empirical results show DSCCLR outperforms 17 other clustering methods.

Conclusions:

• DSCCLR offers a scalable and generalizable deep clustering solution.
• The proposed method effectively captures intrinsic data relationships and manifold structures.
• DSCCLR provides a computationally efficient alternative for spectral clustering on large-scale datasets.