Dynamic-static graph-based ranking contrastive learning for multivariate time series classification

Current representation learning-based multivariate time series classification models typically only consider temporal features and fail to model hidden relationships between different variables. Therefore, this paper proposes a multivariate time series classification model based on dynamic-static graph ranking contrastive learning. The model represents the features of different dimensional variables in multivariate time series as nodes and the latent relationships between variables as edges. While modeling the hidden relationships between multivariate variables, it captures the dynamic variation characteristics of variables through dynamic graph representation. Additionally, it utilizes discrete wavelet transform and its inverse transform to obtain frequency domain features of time series and enhances the data to generate strong and weak views. Meanwhile, by employing node-graph contrastive loss and graph ranking contrastive loss, the model learns robust and discriminative graph representations. This study compares the proposed method with current mainstream multivariate time series classification methods as baselines through comparative experiments on 30 publicly available datasets from UEA. Experimental results demonstrate that the proposed classification approach significantly outperforms existing methods.