ShapeletMIL: A multi-instance learning framework for interpretable multivariate time series classification

Multivariate time series classification (MTSC) plays a crucial role in many real-world domains. However, existing methods often struggle to achieve both high accuracy and interpretability, especially when dealing with complex temporal dynamics and multivariate dependencies. In this paper, we propose ShapeletMIL, a novel framework that integrates shapelet-based representation learning with multi-instance learning (MIL) principles. Our approach learns a set of discriminative and interpretable shape lets in a differentiable manner, enabling the model to identify meaningful sub-sequences that are highly relevant to classification tasks. We further introduce an attention-based pooling mechanism to aggregate instance-level evidence, capturing the importance of each local pattern within a time series. The final prediction is derived from a dual-branch classifier composed of a global feature extractor and a shapelet-matching auxiliary classifier, enhancing both robustness and explain ability. Extensive experiments on 27 UEA benchmark datasets demonstrate that ShapeletMIL reachs SOTA in terms of classification accuracy.