Hybrid method for system analysis of urban mobility quality based on the integration of deterministic engineering methodologies and graph neural networks

Background. Traditional approaches to assessing urban mobility quality rely on departmental methodologies focused either on vehicular traffic or pedestrian accessibility. This “siloed” perspective makes it difficult to identify areas where higher comfort for one group of road users is achieved at the expense of reduced safety for others. The development of spatial analysis methods using geographic information systems does not fully solve the problem, as they do not always account for the non‐linear nature of interactions between adjacent road network sections. At the same time, the emergence of graph neural networks has enabled the processing of interconnected data structures, yet known applications of such networks in the transport domain remain largely limited to short‐term traffic flow forecasting without integration with engineering serviceability indicators. Aim. To develop a hybrid method for system analysis of urban mobility quality that integrates deterministic engineering methods for calculating levels of service with graph neural networks to account for the spatial structure of the road network. Methods. The road network is represented as a graph, the vertices of which correspond to intersections and characteristic points, and the edges correspond to road sections. For each edge, levels of service for all categories of road users (automobiles, pedestrians, public transport, bicycles and personal mobility devices) are calculated using formalized engineering methods. A graph convolutional network is used to propagate information about quality between adjacent sections, taking into account network topology. The network is trained on labeled data with a split of the sample into training (70 %), validation (15 %), and testing (15 %) sets. The testing is carried out on a fragment of the urban road network in the central part of the city, comprising 47 intersections and 89 road sections. Results. The developed method enables scenario modeling of road network sections, taking into account the level of comfort, safety, and the consequences of infrastructure changes. By considering and comparatively analyzing spatial relationships, identifying conflict zones, and the impact of infrastructure facilities on adjacent sections, we achieved a 16 % increase in accuracy (from 0.62 to 0.78), a 15 % increase in recall (from 0.58 to 0.73), and an increase in the F1 score from 0.60 to 0.75 in identifying problem areas. Conclusion. The proposed hybrid method creates a basis for creating intelligent decision support systems in the field of transport planning. It can be used by city administrations to justify priorities for infrastructure development, as well as by design organizations for scenario modeling of the consequences of decisions made. Further research directions include integration of temporal dynamics, application of more complex graph neural network architectures, and scaling of the method to networks of other cities.

Artificial neural network · Flow network · Geographic information system · Graph · Public transport · Serviceability (structure) · Traffic engineering · Traffic and Road Safety · Traffic Prediction and Management Techniques · Transportation Systems and Logistics