ECCA: Efficient Correntropy-Based Clustering Algorithm With Orthogonal Concept Factorization.

Area of Science:

• Machine Learning
• Data Mining
• Unsupervised Learning

Background:

• Clustering large unlabeled datasets efficiently and effectively is a key challenge in unsupervised learning.
• Traditional matrix factorization methods for clustering can be computationally intensive and sensitive to noise and outliers.

Purpose of the Study:

• To propose a novel algorithm, the efficient correntropy-based clustering algorithm (ECCA), for improved clustering efficiency and effectiveness.
• To enhance robustness against noise and outliers in real-world datasets.

Main Methods:

• Developed an orthogonal conceptual factorization (OCF) model to restrict matrix factorization degrees of freedom.
• Introduced an efficient correntropy-based clustering algorithm (ECCA) utilizing an anchor graph for OCF.
• Incorporated correntropy to measure matrix decomposition similarity, enhancing robustness.
• Proposed a novel half-quadratic optimization algorithm for efficient ECCA model optimization.

Main Results:

• ECCA demonstrates significantly improved clustering efficiency, achieving tens to thousands of times speedup compared to state-of-the-art baselines.
• The algorithm shows enhanced robustness and effectiveness on various real-world and noisy datasets.
• Anchor graph construction makes ECCA less sensitive to data dimensionality changes.

Conclusions:

• ECCA offers a highly efficient and robust solution for large-scale unsupervised clustering.
• The proposed method effectively handles noise and outliers, outperforming existing approaches in both speed and accuracy.