Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data

• DOI: 10.1609/aaai.v38i14.29500
• Publication Date: 2024-03-25T11:25:16Z
• Authors: Yucheng Wang; Yuecong Xu; Jianfei Yang; Min Wu; Xiaoli Li; Lihua Xie; Zhenghua Chen
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

Multivariate Time-Series (MTS) data is crucial in various application fields. With its sequential and multi-source (multiple sensors) properties, MTS inherently exhibits Spatial-Temporal (ST) dependencies, involving temporal correlations between timestamps spatial sensors each timestamp. To effectively leverage this information, Graph Neural Network-based methods (GNNs) have been widely adopted. However, existing approaches separately capture dependency fail to the Different sEnsors at Timestamps (DEDT). Overlooking such hinders comprehensive modelling of ST dependencies within data, thus restricting GNNs from learning effective representations. address limitation, we propose a novel method called Fully-Connected Network (FC-STGNN), including two key components namely FC graph construction convolution. For construction, design decay connect across all based on their distances, enabling us fully model by considering DEDT. Further, devise convolution with moving-pooling GNN layer for Extensive experiments show effectiveness FC-STGNN multiple datasets compared SOTA methods. The code available https://github.com/Frank-Wang-oss/FCSTGNN.