TSInsight: A Local-Global Attribution Framework for Interpretability in Time Series Data

Abstract

With the rise in the employment of deep learning methods in safety-critical scenarios, interpretability is more essential than ever before. Although many different directions regarding interpretability have been explored for visual modalities, time series data has been neglected, with only a handful of methods tested due to their poor intelligibility. We approach the problem of interpretability in a novel way by proposing TSInsight, where we attach an auto-encoder to the classifier with a sparsity-inducing norm on its output and fine-tune it based on the gradients from the classifier and a reconstruction penalty. TSInsight learns to preserve features that are important for prediction by the classifier and suppresses those that are irrelevant, i.e., serves as a feature attribution method to boost the interpretability. In contrast to most other attribution frameworks, TSInsight is capable of generating both instance-based and model-based explanations. We evaluated TSInsight along with nine other commonly used attribution methods on eight different time series datasets to validate its efficacy. The evaluation results show that TSInsight naturally achieves output space contraction; therefore, it is an effective tool for the interpretability of deep time series models.

TSInterpret: A unified framework for time series interpretability

“…The time component impedes the usage of existing methods (Ismail et al (2020)). Thus, increasing effort is put into adapting existing methods to time series (e.g., LEFTIST based on SHAP / Lime (Guilleme et al (2019)), Temporal Saliency Rescaling for Saliency Methods (Ismail et al (2020)), Counterfactuals (Ates et al (2021); Delaney et al (2021))), and developing new methods specifically for time series interpretability (e.g., TSInsight based on autoencoders (Siddiqui et al (2021)), TSViz for interpreting CNN (Siddiqui et al (2019))). For a survey of time series interpretability, please refer to Rojat et al (2021).…”

“…Various interpretability methods for time series classification are available (Rojat et al (2021)). However, the usage of those methods is not yet standardized: The proposed methods often lack a) open code (e.g., Siddiqui et al (2021)), b) an easy-to-use interface (e.g., Ismail et al (2020)), or c) visualization (e.g., Guilleme et al (2019)), making the application of those methods inconvenient and thereby hindering the usage of deep learning methods on safety-critical scenarios.…”

Time Series Model Attribution Visualizations as Explanations

“…Siddiqui et al [30] propose with TSinsight an attribution technique but also use multiple line plots to show the attribution as a line plot. Thus, they move the visualization towards small multiples in the form of multiple line plots.…”