XCM: An Explainable Convolutional Neural Network for Multivariate Time Series Classification

Multivariate Time Series (MTS) classification has gained importance over the past decade with increase in number of temporal datasets multiple domains. The current state-of-the-art MTS classifier is a heavyweight deep learning approach, which outperforms second-best only on large datasets. Moreover, this approach cannot provide faithful explanations as it relies post hoc model-agnostic explainability methods, could prevent its use numerous applications. In paper, we present XCM, an eXplainable Convolutional neural network for classification. XCM new compact convolutional extracts information relative to observed variables and time directly from input data. Thus, architecture enables good generalization ability both small datasets, while allowing full exploitation model-specific method (Gradient-weighted Class Activation Mapping) by precisely identifying timestamps data that are important predictions. We first show classifiers public UEA Then, illustrate how reconciles performance synthetic dataset more precise identification regions predictions compared also providing explainability. Finally, can outperform most accurate algorithm real-world application enhancing informative explanations.