SCPaT Improves Time Series Forecasting with Semantic Partitioning.

Multivariate time series forecasting (MTSF) is crucial across many applications, but current patch-based methods have limitations. Fixed partitioning can disrupt meaningful temporal boundaries, multi-scale approaches may create redundant representations, and extendable partitioning often lacks explicit mechanisms for organizing semantic structure or modeling interactions between diverse temporal patterns. To address these issues, the paper proposes SCPaT, a Transformer-based framework built on semantic structured partitioning. SCPaT first adaptively generates semantically consistent units from input sequences. It then constructs a dynamic semantic graph to model dependencies and organize these units into higher-order semantic blocks. Finally, an importance-aware routing mechanism dispatches different semantic blocks to specialized "experts" for customized modeling. Extensive experiments on 12 real-world datasets demonstrate SCPaT's superior effectiveness.