AI-Empowered Spatiotemporal Mobility Intelligence

2023 - Present | UCLA Mobility Lab

Project Overview

This direction develops a unified mobility intelligence framework for generating realistic daily activity-location behavior from heterogeneous data. The direction combines foundation-model learning, household coordination synthesis, and sparse-data augmentation into one coherent modeling agenda.

Motivation

Traditional travel-demand workflows are difficult to transfer across cities and often underrepresent populations with incomplete observations. Existing approaches also tend to separate spatial-temporal prediction from semantic intent and household coordination, which reduces behavioral realism.

Goal

Build transferable mobility models that can infer who travels, when they travel, where they go, and why they move, while maintaining robust performance in data-limited settings.

Method Storyline

Step 1: Learn a universal mobility representation

A foundation model is trained to fuse mobility traces, survey signals, semantic context, and regional structure. This creates transferable latent mobility priors that can be adapted to new geographies.

Foundation model architecture for universal mobility patterns — Foundation model view for cross-domain mobility representation learning.

Step 2: Generate coordinated household behavior at city scale

The learned priors are integrated into a next-generation demand pipeline that models household-level coordination and links behavior synthesis with network simulation.

Household coordinated travel demand generation framework — End-to-end framework from population synthesis to generated demand and simulation input.

Step 3: Address underrepresented populations

A dedicated augmentation stream reconstructs activity chains for shift workers and other hard-to-survey populations, reducing bias in downstream planning analyses.

Shift worker distribution comparison — Distribution-level alignment for underrepresented shift-worker mobility behavior.

Results

OD matrix reconstruction reaches cosine similarity around 0.97 versus benchmark demand outputs.
Daily network VMT alignment reaches JSD around 0.006.
Corridor speed and volume validation reports MAPE around 6.11% in large-scale evaluation.
Shift-worker synthesis quality reports average JSD below 0.02.

Impact

This direction enables a practical path from limited data to policy-grade synthetic demand. It improves inclusion of underserved traveler groups, accelerates scenario setup, and increases confidence when linking AI-generated behavior to operational transportation simulation.