PerCE: Hierarchical Perturbation-Based Counterfactual Explanations for Multivariate Time Series Classification

Betül Bayrak, Kerstin Bach

Abstract: Counterfactual explanations provide intuitive insights into AI model decisions by showing minimal changes needed to alter predictions. However, generating plausible counterfactuals for multivariate time series remains challenging due to complex temporal dependencies, inter-channel correlations, and the need to preserve physiologically realistic patterns, particularly in safety-critical domains like healthcare. This paper presents a novel framework for generating plausible counterfactual explanations for multivariate time series classification, demonstrated on 12-lead ECG data. Our approach employs a hierarchical perturbation strategy guided by permutation-based feature importance, operating at both segment and channel levels to handle temporal dependencies and channel correlations inherent in time series. By anchoring counterfactuals to existing instances from the target class, we enrich plausibility while maintaining minimal modifications. Our evaluation demonstrates significant improvements over the baseline method: 98% validity (vs. 65%), 43% reduction in required modifications, and 75% improvement in proximity to original instances. This work contributes: i) a perturbation-based explanation approach combining instance-based and importance-guided generation for time series, ii) comprehensive evaluation metrics tailored for multivariate time series, and iii) empirical validation on medical data where interpretability is crucial for clinical adoption.

External IDs:dblp:journals/access/BayrakB25