DySA-TGN: Dynamic Self-adaptive Temporal Graph Neural Network for Multivariate Time Series Classification

Multivariate Time Series Classification (MTSC) holds significant importance in various areas. Temporal Graph Neural Networks (TGNNs) that construct dynamic graphs by treating each variate as a node show great potential in MTSC. However, most TGNNs still require a pre-defined graph structure in each time slot and an additional sequence learning module. TodyNet which can learn dynamic graphs without pre-defined graph structure and separate sequence learning was the first TGNN model applied to MTSC. However, TodyNet faces two issues. First, dynamic graph structure learning is purely driven by classification tasks, without fully exploiting dynamic feature representation. Second, TodyNet only aggregates spatial and short-term temporal information without long-term information during information propagation. Based on these issues, we propose the Dynamic Self-Adaptive Temporal Graph Neural Network (DySA-TGN), featuring two core components: 1) The Dynamic Graph Construction based on multi-view feature fusing (DGC-MVFF) module dynamically constructs adjacency matrices by fusing time-varying and time-invariant features; 2) The Graph Isomorphism Network with temporal aggregation (GIN-TA) module aggregates spatial, time-decay and time-variation information to propagate information. Experimental results demonstrated a 5% performance improvement of DySA-TGN over state-of-the-art deep learning methods on 24 UEA benchmark datasets. DySA-TGN significantly enhanced classification accuracy and showed good generalization.