Adaptive bias-aware spatio-temporal graph neural network for time series classification with missing values

Multivariate time series classification is critically important in a variety of fields, including economics, healthcare, and human activity recognition. Recently, numerous classification methods have been proposed for fully observed time series. In practice, however, time series data is often incomplete due to issues such as sensor malfunctions or the high expense of data collection. To overcome this limitation, recent approaches have attempted to exploit the informativeness of missing patterns to enhance temporal or spatial correlation modeling. Nevertheless, these methods typically regard missing patterns merely as binary indicators to distinguish observed from unobserved values, without explicitly modeling their structural properties, limiting the model's adaptability to diverse missing scenarios. To this end, we propose a novel classification framework that explicitly captures local structural information and global authority information from missing patterns. These informative components are integrated into inter- and intra-channel correlation modeling modules by constructing biases, thereby enabling more effective and accurate information propagation along both spatial and temporal dimensions. Experimental results demonstrate that our method consistently outperforms state-of-the-art baselines across three missing scenarios, achieving up to a 20.32 % improvement in classification accuracy.