GPU-Based Multilevel Clustering.

Area of Science:

• Computer Science
• Computational Mathematics

Background:

• Graphics Processing Units (GPUs) offer increasing parallel processing power.
• Limited existing research utilizes GPUs for mesh clustering.
• Mesh clustering requires efficient algorithms to handle large datasets.

Purpose of the Study:

• Introduce a novel multilevel clustering technique implemented solely on GPUs.
• Generalize the mesh-based clustering to broader data clustering applications.
• Address the challenge of missing topological information in general data clustering.

Main Methods:

• Developed a parallel multilevel clustering algorithm optimized for GPU architecture.
• Leveraged spatial coherence for efficient cluster optimization and hierarchical merging.
• Introduced a generalized data clustering approach, including a Local Neighbors k-means algorithm.

Main Results:

• Achieved fast, high-quality, and complete clustering analysis for meshes.
• Demonstrated successful generalization to data clustering with comparable or superior quality to classical methods.
• Showcased excellent scalability, primarily limited by graphics memory capacity.

Conclusions:

• The proposed GPU-based multilevel clustering is efficient and effective for both mesh and general data.
• The generalized approach overcomes limitations of traditional data clustering methods.
• This technique represents a significant advancement in parallel clustering algorithms.