Developing and microsimulating demographic dynamics for an integrated urban model: a comparison between logistic regression and machine learning techniques

Abstract

Studies have shown that sociodemographic attributes significantly influence individuals' transportation choices. However, not all travel demand models do not account for this effect when predicting future travel scenarios. On the other hand, current integrated urban models (IUMs) that incorporate demographic dynamics mostly rely on conventional logit models and rule-based models. These models may not be optimal for complex modeling since they do not fully capture the non-linear relationship between inputs and output. In this research, we explore the feasibility of utilizing machine learning (ML) models to enhance the prediction of demographic dynamics within our proposed IUM—known as ‘STELARS’, in conjunction with conventional logit models. To address the challenge of the black-box nature of ML, we employ an explainable AI technique (xAI) to gain insights into the influence of the factors and compare them with the interpretation revealed by the logit models. Three demographic components are considered: marriage/common-law formation, separation and divorce, and childbirth events, while other components were developed using rate-based models. The results (on the testing dataset) indicate that ML models outperform conventional logit models in terms of overall accuracy by a margin of up-to 3%. However, when considering the true positive accuracy (correctly predicting the event of interest), a significant improvement of 30–48% is observed. Additionally, the xAI analysis reveals consistent interpretation with the logit model. Subsequently, we implemented our demographic dynamics module within our integrated urban modeling system to predict population changes in the Okanagan region of Canada. The multi-year validation of the simulation results against Census data suggests a reasonably close prediction of the observed population. We also optimize the runtime of the demographic dynamics module using vectorization, reducing the simulation time for the demographic changes in our study area (comprising approximately 200,000 individuals living in 85,000 households) to just about 100 s for the total 10 years of simulation. The development and implementation of this advanced demographic dynamics module to accurately predict the life events of individuals adds a fundamental capacity to the STELARS to be built as an event-based microsimulation model.