Irregular Time-Series Modeling

Irregular time-series modeling studies how to forecast, classify, reconstruct, and generate sequences when observations are unevenly spaced, asynchronous, or missing—a common feature of clinical, scientific, and sensor data. Approaches include continuous-time neural ODEs, SDEs, and controlled differential equations; event-based and attention models; and probabilistic methods such as Gaussian processes, normalizing flows, and diffusion models. A central challenge is deciding how to represent observation times and missingness, which may carry predictive information but can also create shortcuts when sampling patterns change. Benchmarks and metrics increasingly test robustness to sparsity and sampling shifts, and results suggest no single modeling strategy consistently works best across settings.