Abstract :
[en] Urban transportation networks are the backbone of modern cities, yet they remain highly vulnerable to disruptions such as accidents, road closures, construction, or flooding. A single failure in one part of the network can propagate system-wide, causing substantial delays, accessibility issues, and cascading effects that impact urban mobility. Identifying which road segments are truly critical—those whose disruption would lead to disproportionate network-level impact—is therefore essential for resilient and intelligent traffic management.
Conventional approaches to criticality assessment rely heavily on static graph-based metrics (e.g., betweenness centrality) or localized simulation scenarios. While informative, these methods lack behavioral realism, generalizability, and scalability. They also define criticality in advance, treating it as a fixed index or rule-based output. This thesis challenges that assumption by introducing a new paradigm: modeling criticality as a learnable property—a dynamic outcome that emerges from how links behave in practice and under stress, and that can be predicted using data-driven methods.
To support this paradigm, the thesis develops a modular pipeline that integrates spatial features, behavioral movement patterns, and simulation-based indicators into supervised machine learning models. The learning process is informed by traffic simulations, where criticality is approximated through the difference in total trip time after link removal. This allows the model to learn from system-wide responses to disruption, rather than relying on predefined thresholds. The result is a scalable and interpretable method that generalizes across cities and disruption scenarios.
The thesis follows a three-article structure. The first contribution, PEMAP, proposes a conceptual and modular framework for post-event traffic management in urban networks. It is composed of four successive phases: (1) indices assessment and evaluation, (2) criticality index calculation, (3) rerouting decision making, and (4) decision application via vehicular networks. While the second phase—criticality modeling—is explored in depth, the work also engages with Phase 1 through a structured review of existing indices and tools, which informed the methodological design. PEMAP thus serves as a structural foundation that frames the broader vision of disruption-aware traffic resilience, motivating scalable, data-integrated methods that go beyond static metrics and support intelligent decision-making under stress. The second contribution introduces VeTraSPM, a trajectory mining model tailored to vehicle movement data, which accounts for directionality, road connectivity, and pattern repetition to generate new behavioral indices. SAMO then comes in to validate the integration of these indices—support, confidence, and sequential impact—within machine learning models for critical link prediction. The final contribution, SMaL-CLIP, presents a scalable machine learning pipeline trained on hybrid features—including graph metrics, simulation outputs (e.g., flow, speed, delay), and trajectory-derived scores—to predict link criticality in two large-scale simulated urban networks (Luxembourg and Monaco).
Results show that the proposed approach achieves high predictive performance, even when trained on limited data, and can generalize across different urban contexts. The combination of simulation-informed labels, interpretable features, and scalable modeling contributes to a new generation of learning-based tools for disruption-aware traffic planning.
In conclusion, this thesis demonstrates that criticality is not a fixed attribute, but a learnable, context-sensitive property. By treating critical links as emergent outcomes of structural context, mobility behavior, and disruption impact, the proposed methodology lays a foundation for scalable, adaptable, and explainable prediction in urban traffic networks.
While this work focuses on the second phase of the PEMAP framework—criticality modeling—it serves as a first stepping stone toward operationalizing the broader vision of disruption-aware traffic resilience. Future research can build on this foundation by implementing the remaining PEMAP phases, including intelligent rerouting strategies and real-time decision dissemination. Additional extensions include validating the SMaL-CLIP pipeline using real-world trajectory data, exploring transferability across diverse urban environments, and embedding these predictive models within real-time decision-support systems for disruption-aware urban mobility planning.
