Graphs and Data in Passenger Transport Systems
Speaker: Oded Cats
Location: Building 36, 01.150 Lipkenszaal
Time: 11:00-11:45
Zoom: link
Abstract: Networks play a critical role in how we conceptualise and solve planning and operations problems in the passenger transport systems domain. Graph representations are increasingly enhanced by supply- and demand-related data which enable to more realistically and dynamically account for supply-demand interactions, for example in relation to passenger delay propagation. In this talk, I will introduce key relevant concepts and graph representations and will illustrate their relevance using a series of selected applications ranging in their temporal and spatial scales.
Uncertainty-aware Probabilistic Travel Demand Prediction with GNN and VAE for Mobility-on-demand Services
Speaker: Tao Peng
Location: Building 36, 01.150 Lipkenszaal
Time: 11:45-12:05
Zoom: link
Abstract: Accurate demand prediction is crucial for efficient Mobility-on-Demand (MoD) system management. While probabilistic forecasting better accounts for uncertainty than traditional point predictions, existing methods often rely on restrictive assumptions or are computationally intensive. We propose STGCN-VAE, a novel deep learning framework combining Spatial-Temporal Graph Convolutional Networks, Variational Autoencoders, and Kernel Density Estimation to capture complex spatial-temporal patterns and quantify uncertainty without strong parametric assumptions. Experiments on real-world MoD datasets show that STGCN-VAE outperforms state-of-the-art baselines in probabilistic demand prediction.
Bicycle Travel Time Estimation via Dual Graph-Based Neural Networks
Speaker: Ting Gao
Location: Building 36, 01.150 Lipkenszaal
Time: 12:05-12:25
Zoom: link
Abstract: In many cities, cycling is becoming a key part of sustainable urban transport. Yet, while car travel times have been widely studied, bicycle travel time estimation remains underexplored—despite its growing importance for traffic management, route planning, and policy development. This talk addresses that gap. I’ll highlight three key challenges: limited and biased cycling data, the complexity of cyclist behavior, and the lack of well-defined cycling infrastructure. To tackle these, we present DG4b, a Dual Graph-based deep learning approach designed specifically for bicycles. It uses one graph to represent the shared road network and another to model individual trip patterns, combining them to estimate segment speeds and full trip durations. Applied to GPS data from Berlin, DG4b significantly outperforms baseline models in accuracy, while keeping computational complexity low. This work offers new insights into bicycle traffic modeling and opens up promising directions for future research in urban mobility.