Self-Constrained Clustering Ensemble.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Existing clustering ensemble methods often fuse base clusterings simultaneously, limiting the ability to assess individual clustering quality and leverage prior knowledge.
• This one-shot fusion approach restricts adaptability to diverse data distributions and overall performance in unsupervised learning.

Purpose of the Study:

• To propose a novel Self-Constrained Clustering Ensemble (SCCE) algorithm to overcome limitations of existing methods.
• To enhance the adaptability and performance of clustering ensembles by distinguishing base clustering quality and exploiting latent prior knowledge.

Main Methods:

• SCCE utilizes pseudo-labels from current clustering results as self-supervised signals.
• Metric learning is employed to achieve a linear transformation that enlarges inter-class distances and compresses intra-class distances.
• Base clusterings are reclustered in the new metric space, followed by iterative ensemble updating in a self-driven closed loop.

Main Results:

• Theoretical analysis indicates efficient convergence via alternating optimization with comparable computational complexity to mainstream methods.
• Experiments on public datasets show SCCE significantly outperforms representative clustering ensemble approaches.
• The algorithm demonstrates effectiveness and robustness in unsupervised learning scenarios.

Conclusions:

• The proposed SCCE algorithm offers a significant advancement in clustering ensemble methods.
• Its self-supervised and metric learning approach enhances data separability and consistency.
• SCCE provides a robust and effective solution for unsupervised clustering tasks.