Back to search

University of North Texas

Graph Neural Networks for Neighborhood Socioeconomic Modeling from High-Dimensional Census Data

Abstract

dc:description

Socioeconomic prediction at the neighborhood level is essential for understanding inequality, supporting evidence-based policy, and guiding resource allocation in urban and regional planning. Traditional census-based predictive models commonly rely on tabular machine learning methods and often treat census units as independent observations. However, socioeconomic conditions are shaped by spatial proximity, demographic similarity, housing markets, labor access, and other relational neighborhood effects that are not fully captured by feature-only models. This dissertation develops and evaluates a graph-aware framework for neighborhood socioeconomic modeling using high-dimensional census data. The proposed framework integrates representation learning, graph construction, supervised prediction, and unsupervised structure discovery. First, high-dimensional census indicators are cleaned, normalized, and transformed into compact representations using principal component analysis and autoencoder-based Census2Vec embeddings. Second, census units are represented as nodes in a k-nearest-neighbor similarity graph, where edges encode socioeconomic similarity among neighborhoods. This graph structure enables the use of graph neural networks, including Graph Convolutional Networks and GraphSAGE, to incorporate relational context through message passing. Third, hybrid models are developed by combining GNN-derived embeddings with original tabular census features and training downstream machine learning regressors. The framework is evaluated on key socioeconomic indicators, including family median income, median gross rent, per capita income, and poverty percent, using standard regression metrics such as MAE, RMSE, and R². The results show that graph-aware and hybrid modeling approaches can improve socioeconomic prediction by capturing both local feature information and structural neighborhood context. In addition, unsupervised graph-based analysis using Louvain community detection reveals meaningful socioeconomic groupings and supports robustness assessment through modularity, stability, and conductance diagnostics. Overall, this dissertation demonstrates that combining census representation learning with graph neural networks provides a scalable and interpretable direction for socioeconomic modeling. The proposed framework bridges traditional tabular machine learning and modern graph learning, offering practical value for policy analysis, urban planning, and future spatiotemporal forecasting of neighborhood change.

Degree

thesis:*
Grantor dc:publisher
University of North Texas
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Hosseini, Farahnaz 1985-
Contributors dc:contributor
  • Xiao, Ting
  • Albert, Mark V.
  • Bhowmick, Sanjukta
  • Li, Yuan

Subjects

dc:subject × 3

Rights

dc:rights
Statement dc:rights
  • Public
  • Hosseini, Farahnaz 1985-
  • Copyright
  • Copyright is held by the author, unless otherwise noted. All rights Reserved.
Language dc:language
English

Identifiers

dc:identifier.*
Identifier
local-cont-no: submission_4961
ark: ark:/67531/metadc2652467

Chain of custody

source
Harvested from
University of North Texas
Base URL
digital.library.unt.edu/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Hosseini, Farahnaz 1985-. Graph Neural Networks for Neighborhood Socioeconomic Modeling from High-Dimensional Census Data. University of North Texas, 2026. https://digital.library.unt.edu/ark:/67531/metadc2652467/