Objective Evaluation of Deep Visual Interpretations on Time Series Data

The correct interpretation and understanding of deep learning models is essential in many applications. (Explanatory) visual interpretation approaches for image and natural language processing allow domain experts to validate and understand almost any deep learning model. However, they fall short when generalizing to arbitrary time series data that is less intuitive and more diverse. Whether a visualization explains the true reasoning or captures the real features is more difficult to judge. Hence, instead of blind trust we need an objective evaluation to obtain reliable quality metrics. This paper proposes a framework of six orthogonal quality metrics for gradient- or perturbation-based post-hoc visual interpretation methods designed for time series classification and segmentation tasks. This comprehensive set is either based on "human perception" or on "functional properties". An extensive experimental study includes commonly used neural network architectures for time series and nine visual interpretation methods. We evaluate the visual interpretation methods with diverse datasets from the UCR repository as well another complex real-world dataset. We show that none of the existing methods consistently outperforms any of the others on all metrics while some of them are ahead in either functional or human-based metrics. Our results allow experts to make an informed choice of suitable visualization techniques for the model and task at hand.