Route Choice Prediction Through User Behavior Analysis: Towards Robustness Assessment Under External Perturbations

We present a preliminary study of route choice prediction and its robustness to small, realistic perturbations, with emphasis on attack surfaces that matter for transport and energy infrastructure. Using a synthetic, richly annotated route dataset, we train per-user Random Forest models (single-model test accuracy 82 %) and aggregate explanations via model-specific Gini importances and Local Interpretable Model-Agnostic Explanations (LIME) attributions. Feature analysis shows that distance and duration dominate the global Gini ranking while LIME highlights temporal indicators as highly explanatory at the instance level. We probe sensitivity with nearest-neighbor counterfactuals and find that modest edits-e.g., sub-kilometre distance shifts (0.4−1.2 km) or small hierarchy changes-can flip predicted choices; an ensemble majority vote over ten user models generalizes to a held-out user with 80 % accuracy. Building on these findings, we formalize a problem-space threat model, discuss the inverse feature-mapping challenge that an attacker must solve to translate feature goals into map/sensor edits, and consider realistic vectors such as fake charger listings and traffic spoofing. ... mehrWe conclude with mitigation directions and a discussion of regulatory implications under the EU AI Act, and outline next steps including in-vehicle validation.