SurF: A Generative Model for Multivariate Irregular Time Series Forecasting

Proposes SurF, a generative model using the Time Rescaling Theorem as a learnable bijection to map irregular multivariate event sequences to i.i.d. exponential noise, enabling single-model training across heterogeneous event-stream datasets.