CATodyNet: Cross-attention temporal dynamic graph neural network for multivariate time series classification

Multivariate time series classification is widely applicable to finance, healthcare, and meteorology; therefore, it is a valuable data-mining task. However, existing methods rely significantly on strong priors owing to onedimensional temporal convolution, or result in error accumulation owing to the information transmission of time slots during temporal dynamic graph fusion. Therefore, a network known as the cross-attention temporal dynamic graph neural network (CATodyNet) is proposed to address these issues. CATodyNet constructs a new convolution stack module to replace one-dimensional temporal convolution and designs a mixed cross- attention module to learn the hidden relationships between variables. Furthermore, CATodyNet streamlines the existing temporal dynamic graph-fusion method and improves the model's convergence speed and classification accuracy. Experimental results show that, compared with existing methods, CATodyNet achieves the best average classification accuracy on 22 University of East Anglia (UEA) time-series classification public datasets and performs well in some multivariate time series classification tasks with non-alignment, missing values, and multiple domains. CATodyNet is expected to provide better decision support than other methods for multivariate time-series data classification. The code used in this study is available at https://github.com/ HQYWY/CATodyNet.