Hybrid machine learning-based approaches for modeling bikeability

‘Bikeability’ is the quantitative assessment of the aggregate influence of natural and built environment features as barriers or facilitators to bicycling. An emerging field, bikeability research incorporates a diversity of factors and approaches, yielding a variety of results. This variability limits both the generalizability of findings and the practical impacts of this research. This study explores machine learning methods as a pathway toward greater convergence of empirical approaches to bikeability modeling. We disaggregate bikeability indicators into four groups: (1) bicycling infrastructure, (2) safety, (3) ambient environment, and (4) accessibility. To derive bikeability indicator weights, we employ a Negative Binomial Regression (NBR) along with two ensemble machine learning algorithms, Random Forest Regression (RFR), and eXtreme Gradient Boosting Regression (XGBR). We then employ the COmplex PRoportional ASsessment (COPRAS) model to score indicators and consider the influences of both the positive and negative criteria, with Sydney, Australia, as a case study. The resulting bikeability scores were statistically validated using bicycle count survey data. The key bikeability factors identified were: (1) destination accessibility, (2) air quality, (3) bikeway and traffic signal density, and (4) bikeway separation from motor vehicle traffic. Performance of the three hybrid models (COPRAS-NBR, COPRAS-RFR, and COPRAS-XGBR) indicates their capacity to handle the complex relationship between bicycling ridership and bikeability indicators and assess bikeability in a way that could support methodological convergence in the field. Findings suggest place-based interventions have an important role to play in supporting bicycling.