New AI Model Improves Irregular Time Series Forecasting

This research introduces the Debiased Neural Basis-Function Network (DNBNet), a novel approach to tackle the complexities of irregular time series forecasting. Irregular time series, common in fields like healthcare and meteorology, are characterized by sparse observations and non-uniform sampling, making accurate prediction challenging for existing methods. Current approaches often aggregate irregular observations into fixed-dimensional coefficients using predefined basis functions, but this can lead to asymptotic bias and limited adaptability to diverse temporal patterns. DNBNet addresses these issues with a debiased neural basis-function response mechanism. This mechanism corrects bias through importance sampling and parameterizes basis functions with neural networks, allowing them to adapt to various temporal patterns. Additionally, DNBNet incorporates a multi-scale decomposition module and a mass-aware fusion mechanism to capture richer representations from sparse data, along with a dual-branch decoder for forecasting. Extensive experiments on multiple real-world datasets demonstrate DNBNet's superior effectiveness and generalizability across diverse irregular time series scenarios.