Sensitivity-propagated dual-frequency graph neural network for multivariate time series forecasting

Abstract

Graph Neural Networks (GNNs) have become one of the mainstream frameworks in multivariate time series (MTS) forecasting due to their powerful spatio-temporal dependency modeling capability. The process of extracting spatio-temporal features can be summarized into three stages: graph generation, graph convolution, and node updating. However, existing works recognize that the quality of the generated graph significantly impacts model performance, while overlooking that effective node updating can produce richer series representations. Furthermore, existing GNNs exhibit a pronounced bias toward capturing low-frequency temporal patterns, with inadequate attention to high-frequency components. Therefore, we propose SensGCN, a novel dynamic graph spatio-temporal network by introducing the concept of series sensitivity features to optimize the node updating process. Built upon a graph convolutional Gated Recurrent Unit (GRU) framework, SensGCN derives sensitivity features from series volatility patterns under non-autocorrelation conditions. These features subsequently guide node updating after aggregating external series information through graph convolution. Additionally, a novel dynamic graph estimation method is developed that extracts high-frequency components via series decomposition to jointly model time-varying spatial dependencies in MTS data, thereby enhancing GNNs’ capability in learning high-frequency features. Extensive evaluations across five public datasets show that our SensGCN achieves competitive or state-of-the-art performance in both multi-step and single-step forecasting tasks. Notably, in multi-step forecasting with a predefined graph structure, SensGCN achieves the best performance in four out of six cases and consistently attains the lowest MAE, outperforming the best baselines by up to approximately 1.3 %.