Hao Xu

We propose FactoST, a two-stage spatio-temporal foundation model that factorizes universal temporal pretraining from spatio-temporal adaptation, enabling efficient and generalizable forecasting across domains.