Abstract
The forecasting of irregular multivariate time series (IMTS) is a criticaltask in domains like healthcare and climate science. However, this task facestwo significant hurdles: 1) the inherent non-uniformity and missing data inIMTS complicate the modeling of temporal dynamics, and 2) existing methodsoften rely on computationally expensive architectures. To address these dualchallenges, we introduce APN, a general and efficient forecasting framework. Atthe core of APN is a novel Time-Aware Patch Aggregation (TAPA) module thatintroduces an aggregation-based paradigm for adaptive patching, moving beyondthe limitations of fixed-span segmentation and interpolation-based methods.TAPA first learns dynamic temporal boundaries to define data-driven segments.Crucially, instead of resampling or interpolating, it directly computes patchrepresentations via a time-aware weighted aggregation of all raw observations,where weights are determined by each observation's temporal relevance to thesegment. This approach provides two key advantages: it preserves data fidelityby avoiding the introduction of artificial data points and ensures completeinformation coverage by design.The resulting regularized and information-richpatch representations enable the use of a lightweight query module forhistorical context aggregation and a simple MLP for final prediction. Extensiveexperiments on multiple real-world datasets demonstrate that APN establishes anew state-of-the-art, significantly outperforming existing methods in bothprediction accuracy and computational efficiency.