Autoencoder-Based Latent Block-Diagonal Representation for Subspace Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Subspace clustering methods like block-diagonal representation (BDR) typically use shallow linear models.
• Real-world data often exhibits nonlinear structures, which are not accurately captured by existing linear BDR methods.
• This limitation hinders the faithful reflection of intrinsic relationships among samples.

Purpose of the Study:

• To propose a novel latent block-diagonal representation (LBDR) model for subspace clustering.
• To address the limitations of linear models in capturing nonlinear data structures.
• To enhance the accuracy and effectiveness of subspace clustering on complex datasets.

Main Methods:

• The proposed LBDR model jointly learns an autoencoder and a BDR matrix.
• An autoencoder with a nonlinear encoder and linear decoder learns effective features from nonlinear samples.
• These learned features form a new dictionary for a linear model with block-diagonal regularization, suitable for spectral clustering.

Main Results:

• The learned features are theoretically proven to reside in a linear space, validating the self-expression model's effectiveness.
• Extensive experiments on diverse real-world datasets demonstrate the superiority of LBDR.
• LBDR significantly outperforms state-of-the-art subspace clustering approaches.

Conclusions:

• The novel LBDR model effectively performs subspace clustering on nonlinear data structures.
• LBDR offers a robust approach by combining nonlinear feature learning with block-diagonal representation.
• The method shows significant improvements over existing techniques, highlighting its potential for complex data analysis.