DNBNet: A Debiased Neural Basis-Function Network for Irregular Time Series Forecasting
Researchers have introduced the Debiased Neural Basis-Function Network (DNBNet), a new approach to irregular time series forecasting, in a preprint on arXiv. Irregular time series data, common in healthcare and meteorological observation, suffer from sparse observations and non-uniform sampling. Existing methods rely on predefined basis functions to convert irregular observations into fixed-dimensional coefficients, but these face two limitations: a potential non-vanishing asymptotic bias due to ignoring timestamp sampling density, and limited adaptability to diverse temporal patterns. DNBNet aims to overcome both problems by incorporating a debiased neural basis. The paper is available at arXiv:2608.17284.
Key facts
- Irregular time series forecasting is crucial in healthcare and meteorological observation.
- Irregular time series have sparse observations and non-uniform sampling.
- Existing methods aggregate irregular observations into fixed-dimensional estimated response coefficients using predefined basis functions.
- The modeling paradigm ignores the sampling density of timestamps.
- Predefined basis functions have limited adaptability to diverse temporal patterns.
- DNBNet is proposed to address these challenges.
- DNBNet's core is a debiased neural basis.
- The research is published as a preprint on arXiv with identifier 2608.17284.
Entities
—