Data-driven causal reasoning for eco-driving behavior analysis in public transportation systems

Public transportation contributes substantially to greenhouse gas emissions, making improvements in bus fuel efficiency a critical operational objective. Although causal inference is well suited for uncovering cause–effect mechanisms in complex systems, it remains underutilized in transportation research, particularly for understanding eco-driving behavior. This paper addresses this gap by proposing a novel causal reasoning pipeline for analyzing the causal drivers of fuel efficiency from observational driving data. The main methodological contribution is a split-and-ensemble extension of the DAGMA-DCE algorithm, which improves the robustness and scalability of causal discovery through structured subsampling and rule-based aggregation. Building on this extension, the proposed pipeline integrates ensemble-based DAG learning, permutation-based falsification testing, and intervention-based root-cause attribution, forming an end-to-end framework that moves beyond correlation analysis toward interpretable causal reasoning. Experiments on synthetic datasets demonstrate that the proposed ensemble approach achieves competitive performance relative to representative gradient-based causal discovery methods and consistently outperforms majority-voting ensembles in terms of robustness. The framework is further applied to a real-world transportation dataset comprising 4006 bus trips with 50 features collected in North Jutland, Denmark. The results identify average speed as a direct causal factor of fuel efficiency at the aggregate level, while subgroup analyses reveal sparse direct effects but multiple stable causal relationships among driving-behavior variables. Root-cause attribution highlights speed regulation, gear selection, and accelerator-pedal control as key contributors to fuel-efficiency variation, emphasizing the role of indirect behavioral pathways. Overall, this study contributes a general, robust, and interpretable causal reasoning framework and demonstrates its applicability for eco-driving analysis and energy-efficient public transportation operations.