Integrating Ensemble Machine Learning and Structural Equation Modeling to Predict and Explain Biking Demand in Kigali: A Data-driven Approach for Sustainable Urban Mobility

Biking-both shared and non-shared-has become a vital component of sustainable urban mobility across African cities like Kigali. Despite this progress, empirical research and demand modeling of biking behavior remain limited. This study predicts and explains biking behavior in Kigali by integrating advanced ensemble machine learning (ML) techniques with structural equation modeling (SEM).

A dataset of 6,386 observations was compiled by merging survey responses on biking with secondary data on weather and air quality. Both traditional statistical and advanced ensemble models were developed for comparison. The dataset was partitioned into training (70%) and testing (30%) subsets, with correlation-based and model-based feature selection applied. SEM examined latent constructs representing spatial, social-demographics, temporal, environmental, and attitudinal factors.

Ensemble ML models substantially outperformed traditional approaches, with random forest and XGB classifiers achieving the highest predictive performance. The SEM demonstrated good model fit and explained the variance in biking frequency. Perceived station accessibility emerged as the strongest determinant of biking behavior, while temporal and environmental factors indirectly influenced demand patterns.

The combination of ML and SEM has revealed a coexistence of accurate prediction and behavioral insight. Accessibility emerged as central to biking uptake, highlighting the potential of station placement and spatial equity. Indirect effects of temporal and environmental conditions highlight the impact of user perceptions in shaping biking demand.

Integrating ensemble ML and SEM provides predictive robustness and behavioral insight. The findings highlight that improving spatial accessibility and adopting adaptive urban planning strategies enhance sustainable biking uptake.