Stefan Huber
I am Stefan, a versatile scientist with expertise at the intersection of transport geography (Diploma) and traffic engineering (PhD). Since 2012, I have dedicated my research, teaching, and consulting efforts to the analysis and modeling of traffic behavior. With a solid foundation in both fields—geography and transport planning—I often bring a unique perspective, leading to different questions and explanations. In my doctoral thesis, I analyzed bicycle route choice preferences using GPS data. I currently co-chair the bicycle research group BIKELAB and chair the subgroup ‘Analysis and Modeling of Active Transport,’ with a particular emphasis on cycling. However, cycling is not just a field of study for me—it is also my passion, whether on an MTB, racing bike, or city bike.
Sessione
Analysing bicycle route choice is highly relevant for transport planning and modelling, as it directly influences the design and optimization of urban transportation systems. Numerous studies have investigated bicycle route choice, providing insights into cyclists' preferences and behaviours. Although machine learning methods are increasingly applied in transport planning, there remains a notable gap in research focusing on their use in analysing and modelling bicycle route choice.
This study aims to assess the utility of machine learning approaches in comparison to established methods such as the multinomial logistic regression (MNL) model. Therefore, a comprehensive GPS dataset comprising 25,730 GPS tracks from 1,361 cyclists in Dresden, Germany, was used. The data was processed and enriched with network data to incorporate relevant factors influencing route choice (e.g., cycling infrastructure, slope, surface, safety, land use etc.). Subsequently, the processed data was used to (a) estimate an MNL model and (b) train various machine learning models, including support vector machines, decision trees, random forests, and neural networks. Furthermore, the effect strength and impact direction of independent variables, as well as overall model accuracy, were determined and compared across all models.
The results reveal that machine learning methods perform comparably well to the traditional MNL model. However, interpreting the results remains challenging. The normalized feature importance of ML models indicates similar relevance for different independent variables as the MNL coefficients. Additionally, the partial dependence of each variable generally suggests the same direction of impact (positive or negative). Combining both measures in a directed normalized feature importance suggests that ML models capture the influence of independent variables on route choice probability similarly to the MNL model. However, while ML models slightly outperform the MNL model in terms of accuracy (86%), their performance varies, ranging from 83% (random forest) to 89% (support vector machine).