Author Type

Graduate Student

Date of Award

Fall 11-24-2025

Document Type

Thesis

Publication Status

Version of Record

Submission Date

November 2025

Department

Computer and Electrical Engineering and Computer Science

Degree Name

Master of Science (MS)

Thesis/Dissertation Advisor [Chair]

KwangSoo Yang

Abstract

This thesis presents a novel approach to node-attributed spatial graph partitioning (NSGP) that addresses limitations in existing homogeneity measures. We introduce a Kullback-Leibler (KL) divergence-based formulation we name KLD that measures attribute similarity by comparing each node’s distribution to its subgraph centroid, providing a computationally efficient method of measuring homogeneity. Integrated within a multilevel graph partitioning framework and Fiduccia-Mattheyses refinement, our approach achieves approximately 90% runtime reduction compared to the baseline Clustering and Local Refinement (CLR) algorithm while maintaining comparable partition quality. Applied to COVID-19 infection data combined with American Community Survey demographics across U.S. FIPS codes, KLD successfully identifies spatially contiguous, demographically homogeneous communities. Experimental results on graphs ranging from hundreds to thousands of nodes demonstrate favorable scaling behavior, making the approach practical for large-scale spatial analysis where both geographic proximity and attribute similarity must be jointly optimized.

Share

COinS