Irregularly sampled sensor data are ubiquitous in healthcare, IoT, and cybersecurity, yet most deep models assume uniform sampling. **IA-GNN** tackles this by: * Constructing a **temporal-spatial graph** where nodes are instances (irregular timestamps) and edges capture multi-scale temporal proximity. * Applying **instance-level attention** to weight informative timestamps, improving interpretability over standard sequence models. * Leveraging **self-supervised pre-training** to mitigate sparsity and label imbalance. On five real-world datasets, IA-GNN improves F1 by up to **7 pp** over strong baselines (T-LSTM, Time2Vec-GNN) while offering saliency maps that highlight clinically relevant events.Recommended citation: **Kun Han**, Abigail M.Y. Koay, Ryan K.L. Ko, Weitong Chen & Miao Xu (2025). “Adapting to the Stream: An Instance-Attention GNN Method for Irregular Multivariate Time Series Data.” *Frontiers of Computer Science*, 19(8):198340. https://doi.org/10.1007/s11704-024-40449-z
Download Paper | Download Slides
