Breaking Data Silos: Towards Open and Scalable Mobility Foundation Models via Generative Continual Learning

Summary

This paper talks about MoveGCL, a new way to build big AI models that understand human movement patterns across different places without sharing private raw data. It uses a special type of learning called generative continual learning combined with a smart model design to let many data holders work together on training the model while keeping their data private.

What's the problem?

The problem is that human mobility data is very sensitive and spread out in different organizations, making it impossible to just combine all the data in one place for training big AI models. Also, the data is very different depending on location and population, and updating models over time usually causes them to forget important things they learned before. These challenges make it hard to build good, up-to-date mobility models that respect privacy.

What's the solution?

The solution is MoveGCL, which lets each data holder generate synthetic or fake movement data that looks like their real data without revealing it. This synthetic data is replayed during training to help the model remember past information while it learns new patterns. The system also uses a special Mixture-of-Experts Transformer that can adapt to different local mobility patterns by routing inputs through expert modules designed for those patterns. Additionally, a step-by-step adaptation process helps keep the model stable as it updates continuously.

Why it matters?

This matters because it enables many groups, like cities or companies, to collaboratively create strong AI models that understand human movement better without sharing private data. Such models can improve urban planning, traffic management, and transportation systems while protecting people's privacy and ensuring the models keep learning and adapting to new data over time.