Title: TiWeaver: Unified Temporal Dynamics Modeling via Contextual Patching

Abstract: Accurate multivariate time series forecasting is essential for numerous practical applications, ranging from stock market analysis and health monitoring to weather prediction. However, the heterogeneity of data sources often results in time series with varied temporal patterns, frequently complicated by irregularities like missing data points and inconsistent sampling rates. These inconsistencies generate complex, asynchronous temporal relationships across different channels. Consequently, traditional models relying on static patching strategies often struggle to adapt to such diverse datasets, which limits forecasting precision. To address this challenge, we introduce TiWeaver, a comprehensive framework capable of adaptively managing temporal dynamics and fine-grained inter-channel dependencies. Our approach features two key innovations: a Graph-Guided Adaptive Tokenizer (G²AT), which segments time series into patches with high contextual coherence by evaluating both temporal density and representation consistency; and a Fine-grained Asynchronous Dependency Extractor (FADE), which captures detailed asynchronous inter-channel relationships alongside long-term historical dependencies. We validated TiWeaver across 12 real-world datasets, where it demonstrated state-of-the-art performance, surpassing existing methods by as much as 25%. These findings highlight the framework’s robustness and efficacy across various domains and data profiles.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC