← Back to Projects

Next Generation Traffic Demand Model: Deep Learning for Human Mobility Synthesis

This research is conducted under the Hidden Activity Signal and Trajectory Anomaly Characterization (HAYSTAC) project, funded by The Intelligence Advanced Research Projects Activity (IARPA) at the Office of the Director of National Intelligence. HAYSTAC aims to establish models of "normal" human movement across times, locations, and people in order to characterize what makes an activity detectable as anomalous within the expanding corpus of global human trajectory data.

We propose a deep learning model to generate synthetic human mobility data based on socio-demographic information and household characteristics. The model is transferable and can be fine-tuned with local data, making it suitable for data-limited regions. By leveraging HTS data, an effective loss function, and optimizing input construction, even a vanilla transformer model proves highly effective in addressing this complex problem. Additionally, we integrate mobility generation with activity location assignment, validating the model's performance by embedding generative travel demand into a large-scale simulation network.The model can be fine-tuned with local open-source data, allowing it to adapt to and accurately represent mobility patterns across diverse regions.

DeepAct Model Workflow
Figure 1: The DeepAct Model Workflow

Interactive Visualization: DeepAct Model

This visualization (adapted from Flowing Data) demonstrates human activity patterns over time using the DeepAct model. Each dot represents an individual, and colors indicate different activities. The visualization runs through a 24-hour cycle showing how people move between different activities throughout the day. Given the synthetic socio-demographic information and household characteristics of each agent, our model can synthesize their complete activity chain and the location of each activity.

The framework builds on the Multi-Agent Transport Simulation (MATSim) to tackle the challenges induced by urbanization and changing mobility patterns. By loading the synthesized travel demand into the LA roadway network, we can generate the synthetic traffic flow and further evaluate the transportation system performance based on the generated human mobility patterns.

Model Evaluation and Impact

The model is evaluated on a nationwide dataset of the United States, where it demonstrates superior performance in generating activity-location chains that closely follow ground truth distributions. Further tests using state- or city-specific datasets from California, Washington, and Mexico City confirm its transferability.

This innovative approach offers substantial potential to advance mobility modeling research, particularly in generating synthetic human mobility data. This can provide urban planners and policymakers with enhanced tools for simulating mobility in diverse regions and better informing decisions related to transportation, urban development, and public health.