An Interpretable GeoAI Framework for Analyzing Multi-Ethnic Settlement Dynamics: Evidence from the Greater Toronto Area, 2001-2021

Abstract

This dissertation develops an interpretable Geospatial Artificial Intelligence (GeoAI) framework for understanding multi-ethnic settlement patterns in contemporary Canadian cities, demonstrating that computational sophistication need not sacrifice theoretical transparency or democratic accountability. Through three interconnected investigations spanning 52 major Canadian cities, 30,091 dissemination areas, and nine major ethnic groups (China, India, Philippines, Pakistan, Sri Lanka, Iran, Portugal, Italy, United Kingdom) over two decades (2001-2021), this research transforms multi-ethnic settlement from a descriptive sociological phenomenon into a predictive science grounded in interpretable physics-informed models. The research addresses a fundamental tension in urban artificial intelligence: while cities increasingly deploy algorithmic systems to manage complex urban dynamics, the dominant paradigm of black-box optimization systematically fails to achieve stated goals while potentially harming vulnerable communities. A critical analysis of 157 urban AI deployments (2015-2024) reveals pervasive "metrics traps" where impressive technical accuracy, such as ShotSpotter's 97% acoustic precision yielding only 9.1% crime-fighting effectiveness, consistently fails to translate into meaningful social outcomes. This critique establishes the ethical and methodological imperative for interpretable approaches in urban demographic analysis. The core methodological contribution is the development of a graph-based physics-informed neural network (GraphPDE) that embeds multi-ethnic reaction-diffusion dynamics while learning interpretable demographic parameters. Adapting Turing's pattern formation theory to spatial graphs, this framework reveals that ethnic settlement patterns emerge from self-organizing spatiotemporal processes with quantifiable characteristics: ethnic-specific spatial scales ranging from 34.5 km (Philippines) to 63.0 km (United Kingdom); pattern formation regimes segregating groups into spots (UK, Portugal), stripes (China, India, Philippines), and labyrinthine (Iran, Pakistan, Sri Lanka) morphologies; and critical nucleation thresholds varying from 658 individuals (China) to 8,132 individuals (India) for spatial clustering emergence. The learned attention-based interaction mechanism quantifies previously unmeasurable inter-ethnic dynamics, revealing that Chinese populations exhibit strong negative self-competition (-19.8) driving spatial dispersal while maintaining facilitative relationships with multiple groups, Philippines-Pakistan mutual attraction, and United Kingdom-Italy mutual repulsion. The Multi-Ethnic Spatial Mixture of Experts (MESMoE) framework synthesizes physics-informed modeling with machine learning, achieving state-of-the-art predictive performance (R² = 0.80-0.83) while maintaining complete interpretability through regime-specific expert modules for colonization, jump processes, decline, and continuous diffusion dynamics. This architecture demonstrates that incorporating domain knowledge enhances rather than compromises predictive accuracy, with physics-informed components accounting for 57.2% of prediction variance. The framework reveals systematic differences in how ethnic communities respond to urban infrastructure: Chinese and Filipino populations show amplification factors strongly correlated with transit accessibility (r = 0.58 and 0.61), while Indian populations demonstrate stronger correlation with housing variables (r = 0.47). Configuration landscape analysis identifies multiple stable settlement configurations with quantifiable transition barriers, revealing that demographic transitions require sustained interventions over decadal timescales due to asymmetric barriers creating lock-in effects. The temporal evolution of parameters captures non-stationary dynamics across four census periods, with Philippines-origin populations maintaining consistently positive growth rates (0.015-0.023 year⁻¹) while United Kingdom-origin populations transition from positive growth to sustained decline. This research makes four distinctive contributions: (1) developing progressive spatial analysis as a systematic methodology for studying complex urban phenomena through integrated multi-scale investigation; (2) demonstrating interpretable GeoAI that achieves competitive performance without sacrificing transparency; (3) revealing hidden ethnic dynamics through quantification of inter-ethnic interactions, nucleation thresholds, and pattern formation regimes; and (4) providing practical tools bridging academic research and urban planning practice through open-source implementations. The framework transforms abstract geographic concepts into measurable quantities, "sense of place" becomes configuration stability, "community cohesion" translates to interaction strength, and "neighborhood character" maps to position in pattern formation phase space. The policy implications are transformative: understanding asymmetric configuration barriers explains persistent settlement patterns while identifying intervention requirements; regime-specific approaches match policies to demographic dynamics; and cascade effects enable strategic investments generating system-wide benefits. Critical reflection acknowledges fundamental limitations, physics cannot capture individual agency, cultural meaning, or structural inequalities, yet the framework complements rather than replaces other ways of understanding urban dynamics. This dissertation demonstrates that interpretable GeoAI can bridge the persistent divide between quantitative spatial science and critical human geography, proving that mathematical rigor need not sacrifice social awareness. The discovery that Canadian cities' ethnic geography follows learnable physical dynamics while maintaining cultural distinctiveness suggests that diversity and order are complementary aspects of urban organization, offering both cautionary lessons about algorithmic governance and hopeful possibilities for creating more equitable, integrated multicultural cities.