Fast Clustering With Anchor Guidance.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Clustering methods partition data but often suffer from hyper-parameter tuning issues and high computational costs.
• Existing graph clustering approaches face challenges with scalability and model applicability due to complex regularization terms.

Purpose of the Study:

• To propose a novel Fast Clustering model with Anchor Guidance (FCAG) that addresses limitations of existing clustering techniques.
• To develop an efficient and scalable clustering solution suitable for large datasets and out-of-sample extension problems.

Main Methods:

• Developed the Fast Clustering model with Anchor Guidance (FCAG) utilizing bipartite graph properties.
• Implemented and compared three optimization algorithms: Projected Gradient Descent (PGD), Iteratively Re-Weighted (IRW), and Coordinate Descent (CD).

Main Results:

• FCAG effectively avoids trivial solutions without requiring extra regularization terms.
• The Coordinate Descent (CD) optimization method demonstrated superior performance and efficiency in experiments.
• FCAG exhibits better performance and lower time complexity compared to existing bipartite graph models.
• Theoretical and experimental analysis showed PGD converges to IRW as the learning rate approaches infinity.

Conclusions:

• FCAG offers an efficient, scalable, and robust solution for graph clustering, particularly for large-scale problems.
• The Coordinate Descent optimization method is recommended for solving the FCAG model.
• FCAG provides a practical alternative to existing methods, handling out-of-sample extensions effectively.