Diffusion Process-Based Model for Network Trajectory Propagation: Formulation, Data Mining, and Cross-Comparative Analysis

This paper introduces a diffusion process-based approach to address trajectory generation and vehicle propagation within traffic networks. This approach presents dynamic transition probabilities to dictate the navigation of trips through the network. The proposed probabilities incorporate connectiveness and attractiveness of network cells with propagation threshold associated with trip length to effectively replicate real-world observations. This enables trajectory generation that closely adheres to real-world land-use criteria and traffic patterns. The generation of trajectories within the network arises from a purely data-driven methodology, which bypasses the direct use of the city’s network file. Instead, it uses the demand pattern to reconstruct the network through the application of granular Voronoi cells. This approach offers a more refined spatial representation resulting in cells of different sizes as a function of the demand level. Trajectories were generated by fusing datasets of real-world trips, land-use, and network characteristics from Daejeon, South Korea, ensuring a robust calibration and validation process. Compared to a Recurrent Neural Network (RNN) model and a Mobility Markov Chain (MMC) model, our model is superior in terms of cell and link visits counts, complexity similarity with original dataset, correct cell sequence prediction, and Origin-Destination flow pattern that follows the pattern of original trips. The findings underscore the potential of diffusion process-based approaches to enhance transportation planning, showing that non-Artificial Intelligence (non-AI), data-driven methodologies can perform comparably to AI-driven techniques. These methods generate trajectories that closely reflect real-world traffic dynamics while aligning with land-use criteria, offering an explainable, transparent and robust alternative.