MMCont-Net: toward full-scale context modeling for irregular series forecasting

Abstract Forecasting irregularly sampled multivariate time series (IMTS) remains a fundamental challenge in contemporary time series analysis, owing to the inherent non-uniformity of sampling and pronounced inter-variable asynchrony. While recent advances in context learning achieves remarkable performance, microscopic signal distortion and macroscopic spatial isolation issues still limit the effectiveness of irregular series Forecasting. In this paper, we propose MMCont-Net (Micro–Macro Context Modeling Network), MMCont-Net (Micro–Macro Context Modeling Network), a novel framework that pioneers full-scale context modeling. Specifically, MMCont-Net first captures the micro-level intra-patch context within localized signal segments, then models the micro-level cross-patch context across time, and finally reasons over the macro-level inter-variable context across the observations. By this way, MMCont-Net achieves coherent modeling through progressive context integration. Extensive experiments on multiple benchmark datasets, including PhysioNet, Human Activity, and USHCN, demonstrate that MMCont-Net consistently achieves lower forecasting errors while maintaining superior computational efficiency and robustness, demonstrating the importance of the proposed full-scale context modeling paradigm.