Robust Discriminant Subspace Clustering With Adaptive Local Structure Embedding.

Area of Science:

• Data Mining
• Machine Learning
• Pattern Recognition

Background:

• Unsupervised dimension reduction and clustering are often performed separately, potentially misrepresenting subspace cluster structures.
• Existing subspace clustering methods neglect the interplay between low-dimensional representations and input space local structures.

Purpose of the Study:

• To propose a Robust Discriminant Subspace (RDS) clustering model that integrates dimension reduction and clustering.
• To enhance clustering by adaptively embedding local structure information from the original data space.

Main Methods:

• Unified matrix factorization (MF) to integrate dimension reduction and clustering, avoiding extra parameters.
• Constructed a similarity graph to learn local structure, constrained to match low-dimensional representation components.
• Employed the l2,1-norm for robust residual error measurement and efficient optimization.

Main Results:

• The proposed RDS model provides more interpretable clustering outcomes.
• RDS demonstrates superior performance compared to state-of-the-art subspace clustering alternatives on benchmark datasets.

Conclusions:

• The unified approach in RDS effectively addresses limitations of separate dimension reduction and clustering.
• RDS offers a robust and adaptive method for subspace clustering, improving interpretability and accuracy.