Multivariate time series classification (MTSC), which identifies categories of multiple sensor signals recorded in continuous time, is widely used in various fields such as transportation, finance, and medical treatment. The focused challenge remains learning the dependencies between subsequences to capture discriminative patterns while providing convincing explanations. In this paper, we propose a temporal-variable parallel deep learning framework to mine global and local features to achieve a win-win situation in performance and interpretability. Specifically, for harmonizing the inattention blindness of global features, we introduce a graph attention mechanism with global awareness (GAT-g), where the learning of edge representations incorporates both inter-node relationships and the node-to-graph context. Furthermore, for evaluating the feature combinations utility, we exploit game interactions for the first time, which quantifies the utility of feature combination through Shapley values to illustrate the dynamically coordinating representation ability of the model for diverse time series features. In addition, the interpretation module leverages temporal and variable subspace attention distributions to provide instantiated explanations with additive computational complexity, enhancing the comprehension of prediction results. Experimental evaluation on the University of East Anglia (UEA) archive of 30 multivariate time series datasets shows that the proposed method outperforms 12 state-of-the-art methods on 11 datasets.
Keywords: Global awareness; Graph attention networks; Interaction; Interpretation; Multivariate time series classification.
Copyright © 2024 Elsevier Ltd. All rights reserved.