Xin Zhang
San Diego State University, CA·Computer Science
Cities produce constant streams of data about movement: traffic speeds, travel demand, ride-hailing trips, GPS traces. Artificial intelligence models increasingly learn from these streams to forecast conditions and to imitate human travel choices.
- Fields
Computer Science
Engineering
- Topics
large language models
spatial-temporal forecasting
urban computing
- Grants
- 2 known grants, through Aug 2029
What they work on
- Establish diagnostic theory tying spatial-temporal tokenization, cross-modal representation alignment, and distributional drift to the conditions under which LLM-based urban forecasters generalize or break down, producing measurable criteria usable before deployment.
- Develop learning algorithms for LLM-based traffic and travel-demand forecasting that remain faithful to urban constraints and robust across tasks and cities, with guarantees on prediction feasibility under sensor failures and disruptive events.
- Build interactive reasoning and verification mechanisms that let users refine forecasts through human-readable instructions while automatically checking consistency with physical laws, operational requirements, and statistical constraints.
- Create the SIFT framework to erase individual trajectories or behaviors from spatial-temporal imitation learning models through negation trajectory generation and reward shaping, avoiding retraining from scratch.
- Maintain model utility when unlearning records from data-sparse regions using spatial-temporal partitioning and similarity-based augmentation, and supply provable forgetting guarantees via differential privacy on real ride-hailing and transit datasets.
More about the research
Two problems follow. Operators need to trust the forecasts, and the people whose data trained the models need their privacy protected. Both projects work on making mobility AI more dependable, more interpretable, and compliant with data-deletion rules.
Why it matters
Transit agencies, city planners, and ride-hailing and delivery companies increasingly act on AI forecasts of movement. Forecasting tools that signal in advance when they are likely to fail, survive sensor outages and unusual events, and explain which places and moments drove a prediction would make such decisions safer. The unlearning work could let mobility services honor data-deletion requests by removing a person's location history from a deployed model while keeping it useful.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Dongxiao Zhu
Wayne State University, MI·Computer Science
Develop optimization algorithms that make large language model unlearning, removing specific data influences without retraining, more effective, robust and efficient.
Guang Wang
Florida State University, FL·Computer Science
Build cross-domain data curation tools for MobilityNet that combine real mobility datasets from different sources to reduce data bias.
Sanjay Purushotham
University of Maryland-Baltimore County, MD·Computer Science·Early-career
Develop multi-view federated learning algorithms that train shared models across hospitals holding different kinds of patient data, without moving the data between sites.
The school
San Diego State University: 13 labs on PI Pages, $182M a year on research, 142 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.