Learning Decomposed Spatial Relations for Multi-Variate Time-Series Modeling

Modeling multi-variate time-series (MVTS) data is a long-standing research subject and has found wide applications. Recently, there surge of interest in modeling spatial relations between variables as graphs, i.e., first learning one static graph for each dataset then exploiting the structure via neural networks. However, may differ substantially across samples, building all samples inherently limits flexibility severely degrades performance practice. To address this issue, we propose framework fine-grained utilization correlation variables. By analyzing statistical properties real-world datasets, universal decomposition graphs identified. Specifically, hidden can be decomposed into prior part, which applies dynamic varies different necessary to model these relations. better coordinate two relational min-max paradigm that not only regulates common part but also guarantees distinguishability among samples. The experimental results show our proposed outperforms state-of-the-art baseline methods on both forecasting point prediction tasks.