Multivariate time series classification based on spatial-temporal attention dynamic graph neural network

Abstract

Multivariate time series classification (MVTSC) has significant potential for Internet of Things applications. Recently, deep learning (DL) and graph neural network (GNN) methods have been applied to MVTSC tasks. Unfortunately, DL-based methods ignore explicit inter-series correlation modeling. Most existing GNN-based methods treat MVTS data as a static graph spanning the entire temporal trajectory, which inadequately captures changes in inter-series local correlations. To address this problem, we propose the spatial-temporal attention dynamic GNN (STADGNN), which explicitly models dynamic inter-series correlations by constructing the MVTS data into a dynamic graph structure at a finer granularity. It combines discrete Fourier transform (DFT) and discrete wavelet transform (DWT), which extract the global and local features of MVTS data in an end-to-end framework. In dynamic graph learning, spatial-temporal attention mechanisms are employed to simultaneously capture changes in inter-series local correlations and intra-series temporal dependencies without relying on predefined priors. Experimental results on 25 UEA datasets indicate that the STADGNN outperforms existing DL-based and GNN-based baseline models in MVTSC tasks.