Skip to content
Book Open access

PaDi: Patch-wise Decomposition and Latent Interaction for Multivariate Time Series Forecasting

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 34 references

Abstract

Multivariate long-term time series forecasting aims to predict the future trajectories of multiple variables. However, many existing methods oversimplify temporal patches and channel interactions: they encode each raw patch as a single token and model channel relations directly in the original variable space. Such simplification leads to two limitations. At the patch level, raw-patch similarity can diverge from predictive-pattern similarity, since local state shifts may dominate the observed values, even when fluctuation patterns are similar. At the channel level, useful shared dynamics are often concentrated in compact latent structures rather than being uniformly distributed over all pairwise channel interactions. These two limitations reveal patch-level temporal heterogeneity and channel-level structural heterogeneity, suggesting that forecasting models should preserve such structures rather than force them into undifferentiated patch and channel tokens. To address these issues, we propose PaDi, a patch-based forecasting framework that captures such heterogeneity through patch-wise decomposition and latent interaction. Specifically, PaDi disentangles each input patch into base and deviation components and injects a compact base summary into future patch queries as temporal anchors, enabling stable coarse-grained cues to guide the reuse of deviation patterns. It further introduces Latent Channel Attention to exchange cross-variable information through learnable latent channel bases, thereby avoiding redundant dense channel mixing in the original variable space. Extensive experiments on seven public benchmarks under four prediction horizons demonstrate that PaDi achieves superior performance compared with several state-of-the-art baselines, highlighting its effectiveness and stability across diverse forecasting scenarios. The implementation of our model is available at: https://github.com/HYYAlex/PaDi.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.