Reweighted Subspace Clustering Guided by Local and Global Structure Preservation

Area of Science:

• Data Science
• Machine Learning
• High-Dimensional Data Analysis

Background:

• Subspace clustering partitions high-dimensional data into multiple subspaces.
• Current methods focus on similarity matrices and sparse projection matrices.
• Assessing projected subspace dimensionality is challenging, impacting performance with noise or overlap.

Purpose of the Study:

• To propose a novel reweighted subspace clustering (RWSC) model.
• To address challenges in dimensionality assessment and improve clustering performance.
• To enhance robustness and applicability in complex, high-dimensional datasets.

Main Methods:

• Introduced a novel reweighting strategy applied to projected coordinates.
• Developed a reweighted subspace clustering model (RWSC).
• Integrated global scatter structure preservation and adaptive local structure learning.

Main Results:

• Reweighting strategy augments/suppresses coordinate importance, distinguishing overlapping subspaces.
• Redundant coordinates are removed, alleviating bias from imprecise dimensionality.
• RWSC demonstrated improved robustness and applicability on synthetic and real-world datasets.

Conclusions:

• RWSC effectively alleviates bias from imprecise dimensionalities in subspace clustering.
• The model captures intrinsic data structures better through integrated learning.
• Empirical verification confirms the effectiveness and superiority of the RWSC model.