Machine Learning for Investigating Urban Systems: Predicting Human Activities, Business Dynamics, and Electric Vehicle Adoption

Abstract

Cities face unprecedented challenges in adapting to rapid technological and behavioral change, exposing the limitations of traditional transportation and urban modelling approaches. This thesis introduces a novel, data-driven and modular approach for urban analytics centered on explainable machine learning, specifically predicting a person’s activity schedule, business establishment dynamics, and household electric vehicle adoption. Leveraging large-scale, multi-source data from Halifax Regional Municipality, the research develops specialized modules, each independently validated yet engineered for interoperability. The activity scheduling system combines interpretable boosting (EBM, 73.7% accuracy) and deep learning (clustered bidirectional LSTM, Macro F1-score 59.9), achieving robust, equity-focused predictions across diverse demographic segments and capturing nuanced daily activity chains. For business establishment dynamics, a spatial Graph Neural Network is developed to forecast the number of businesses at dissemination-area resolution, achieving reliable predictions with an overall R² = 0.739. Firm-level models predict business sales and employment, revealing how establishment characteristics, accessibility, and economic output interrelate through Explainable Boosting Machines, while also capturing divergences between sales revenue and workforce growth that inform more nuanced transportation planning, particularly under surge scenarios. For household vehicle adoption, interpretable machine learning approaches identify population density, household income, and charging infrastructure as the dominant influences on electric vehicle uptake, with the leading model achieving a strong ROC AUC score of 0.65. These modular, transferable machine learning frameworks offer an evidence-based toolkit for urban policy and scenario analysis. This thesis demonstrates that explainable ML delivers actionable insights for urban planning, paving the way for adaptive, transparent modelling approaches that can succeed and eventually replace lengthy and resource-intensive traditional models as the availability of transportation data expands.