On Consistent Entropy-Regularized k-Means Clustering With Feature Weight Learning: Algorithm and Statistical...

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Real-world data often exhibits clusters within low-dimensional subspaces, posing challenges for traditional clustering algorithms.
• Existing methods for subspace clustering can be computationally intensive and lack theoretical guarantees for large datasets.

Purpose of the Study:

• To develop a computationally efficient and theoretically sound method for learning feature importance in clustering.
• To address the challenge of subspace clustering by integrating an entropy incentive term.

Main Methods:

• Introduced an entropy incentive term within a center-based clustering framework to learn feature importance.
• Developed a scalable block-coordinate descent algorithm with closed-form updates for objective function minimization.
• Utilized Vapnik-Chervonenkis (VC) theory to provide theoretical guarantees, including strong consistency and uniform concentration bounds.

Main Results:

• The proposed method efficiently learns feature importance, enabling effective clustering in low-dimensional subspaces.
• The block-coordinate descent algorithm ensures scalability and computational efficiency.
• Theoretical analysis using VC theory confirms the method's strong consistency and concentration properties.

Conclusions:

• The novel entropy-incentivized clustering approach offers a computationally efficient and theoretically robust solution for subspace clustering.
• Experimental results on benchmark datasets validate the method's effectiveness and scalability.