Diverse Counterfactual Explanations for Anomaly Detection in Time Series
Abstract: Data-driven algorithms for detecting anomalies in times series data are ubiquitous, but
generally unable to provide helpful explanations for the predictions they make. In this
work we propose a post-hoc explainability method that is applicable to any differentiable
anomaly detection algorithm for time series. Our method provides explanations in the
form of a set of diverse counterfactual examples, i.e., multiple perturbed versions of the
original time series that are similar to the latter but not considered anomalous by the
detection algorithm. Those examples are informative on the important features of the time
series and the magnitude of changes that can be made to render it non-anomalous for the
explained algorithm. We call our method counterfactual ensemble explanation, and test it on
two deep-learning-based anomaly detection models. We apply the latter to univariate and
multivariate real-world data sets and assess the quality of our explanations under several
explainability criteria such as Validity, Plausibility, Closeness and Diversity. We show that
our algorithm can produce valuable explanations; moreover, we propose a novel visualization
of our explanations that can convey a richer interpretation of a detection algorithm’s internal
mechanism than existing post-hoc explainability methods. Additionally, we design a sparse
variant of our method to improve the interpretability of our explanation for high-dimensional
time series anomalies. In this setting, our explanation is localized on only a few dimensions
and can therefore be communicated more efficiently to the model’s user.
Submission Length: Long submission (more than 12 pages of main content)
Assigned Action Editor: ~Fuxin_Li1
Submission Number: 457
Loading