LoGCC: Local-to-Global Correlation Clustering for Scalar Field Ensembles

Area of Science:

• Data analysis
• Computational science
• Scientific visualization

Background:

• Correlation clustering (CC) effectively analyzes scalar fields but is computationally intensive.
• Existing CC methods struggle with large datasets and interactive analysis due to high computational cost.

Purpose of the Study:

• Introduce the Local-to-Global Correlation Clustering (LoGCC) framework to accelerate pivot-based CC.
• Leverage spatial structure and correlation transitivity for efficient clustering.
• Enhance scalability and applicability of CC for large-scale and interactive analysis.

Main Methods:

• Developed a two-stage LoGCC framework: a local step using neighborhood graphs and a global step for merging clusters.
• Implemented LoGCC with Pivot and CN-Pivot methods to demonstrate generality.
• Evaluated performance on synthetic, meteorological, and medical image segmentation datasets.

Main Results:

• LoGCC achieved significant speedups: up to 15× for Pivot and 200× for CN-Pivot.
• Demonstrated improved scalability for analyzing larger scalar fields.
• Maintained high cluster quality comparable to existing methods.

Conclusions:

• LoGCC framework substantially accelerates correlation clustering for scalar fields.
• The method enhances scalability, making CC practical for large-scale and interactive applications.
• LoGCC broadens the utility of correlation clustering across scientific domains.