Deep models often trade interpretability for accuracy—especially on multivariate, irregular time-series data. **HITS** (Hierarchical Interpretable Time Series) bridges this gap by: * **Variable-level MIL attention** to spotlight the most informative sensors or channels. * **Patch-level temporal pooling** that captures long-range dependencies without over-segmenting data. * An **information-bottleneck regulariser** ensuring concise, human-interpretable saliency maps. Across 12 healthcare and IoT datasets, HITS matches or exceeds the performance of leading baselines (TimeMIL, PatchTST) while reducing explanation entropy by **32 %**, making insights clearer for domain experts.Recommended citation: Kun Han, Abigail M.Y. Koay, Ryan K.L. Ko, Weitong Chen & Miao Xu (2025). “HITS: Hierarchical Interpretable Time Series Classification via Multiple Instance Learning.” *Proceedings of the International Joint Conference on Neural Networks (IJCNN 2025)*. Rome, Italy.
Download Paper | Download Slides
