Instance-based Counterfactual Explanations for Time Series Classification

1 Introduction

• Most XAI research focuses on image and tabular data, while time-series explanations face multidimensionality, strong feature dependencies, and context-setting challenges.The paper motivates time-series explanations with applications including ECG abnormality detection and food spectroscopy.Source: §1 · pp. 1–2
• Native Guide is a novel, model-agnostic technique that generates counterfactual explanations for time-series classification predictions.The paper frames it as an explainable case-based reasoning solution.Source: §1 · pp. 1–2
• In healthcare, explanations can help practitioners identify informative ECG portions for detecting abnormalities such as myocardial infarction.The paper also points to food spectroscopy and identifying time-series portions relevant to classification.Source: §1 · p. 2

2 Related Work

• Early time-series XAI mainly used visualization and feature-importance methods rather than factual or counterfactual instances.CAM highlights discriminative areas for deep networks; shapelets identify discriminative subsequences, though interpretability concerns have been raised for learned shapelets.Source: §2 · p. 3
• Feature weights can come directly from linear classifiers or be estimated with model-agnostic methods such as LIME and SHAP.Reported stability concerns include perturbation sensitivity; one comparison found LIME poor across evaluated datasets, while saliency methods and SHAP were more robust across architectures.Source: §2 · pp. 3–4
• Time-series prototypes provide class-representative examples but do not identify the most discriminative areas, while existing case-based approaches do not provide counterfactual explanations.An earlier unpublished version retrieved case-base instances but did not integrate discriminative feature information into counterfactual generation.Source: §2 · p. 4

3 Good Counterfactuals for Time Series: Key Properties

• Proximity measures how close a query is to its counterfactual, using distance metrics and a relative counterfactual distance measure.The relative measure compares generated counterfactuals with existing in-sample counterfactual solutions.Source: §3 · pp. 4–5
• Sparse time-series counterfactuals should modify one contiguous, discriminative subsequence rather than scattered individual time points.The paper argues that a simple feature-count sparsity rule from tabular data does not transfer directly to multidimensional time series.Source: §3 · p. 5
• Plausibility may be better assessed by whether a counterfactual falls within or represents the data distribution than by proximity alone.The paper notes that a close counterfactual can still be out of distribution.Source: §3 · p. 5
• The method generates multiple explanations for a query, but diversity should not sacrifice plausibility or semantically meaningful information..Source: §3 · p. 5

• A counterfactual changes query Tq so the black-box classifier predicts an alternative class c′, with proximity, sparsity, plausibility, and diversity prioritized.Source: §4 · p. 6
• Native Guide retrieves an in-sample counterfactual from the case-base, typically the query’s nearest unlike neighbor, which ensures plausibility by staying within the data distribution.Source: §4 · pp. 5–6
• For deep learners, feature weights can direct changes to semantically meaningful features rather than raw time-series values, supporting sparse explanations.Source: §4 · pp. 6–7
• It adapts the native guide by replacing influential contiguous regions of Tq with guide values, using feature-weight magnitude to identify the region.Source: §4 · p. 7
• The modified subsequence can grow iteratively until the classifier predicts c′, reaching the full query-series length in the worst case; other in-sample guides can provide diverse alternatives.Source: §4 · p. 7

5 Testing Native Guide: Two Comparative Experiments

• The tests focused on a fully convolutional neural network classifier and evaluated proximity and sparsity in Experiment 1, then plausibility and diversity in Experiment 2.Experiment 2 used Local Outlier Factor, Isolation Forest, and OC-SVM novelty detection, including OC-SVM on matrix-profile representations.Source: §5 · pp. 7–8, §5.2 · p. 11
• The experiments used five binary and multiclass UCR datasets with archive-provided train-test splits and compared Native Guide against w-CF and NUN-CF.NUN-CF uses the query’s nearest unlike neighbor directly; Native Guide adapts an in-sample counterfactual using CAM feature-importance weights, while w-CF minimizes a loss function.Source: §5 · pp. 7–8
• The authors report that DiCE did not generate diverse counterfactuals within reasonable time limits in these tests.They suggest it may not be well suited to high-dimensional time series, even for shallower ANNs.Source: §5 · p. 8

5.1 Experiment 1: Probing Proximity and Sparsity

• Both Native Guide and w-CF produced explanations significantly closer than existing NUNs on ℓ1 and ℓ2 distances (Wilcoxon, p<0.01).Native Guide constrains perturbations to selected contiguous subsequences, whereas w-CF often changes features at non-contiguous locations.Source: Results and discussion · p. 10
• w-CF’s counterfactuals sometimes contained erratic spikes and non-contiguous changes, while Native Guide edited contiguous subsequences.The authors describe these w-CF spikes as potentially out-of-distribution and note that ℓ∞ can help diagnose erratic feature values.Source: Results and discussion · pp. 10–11, Figure 6 · p. 10
• Native Guide counterfactuals were more proximate than w-CF counterfactuals on ℓ∞, which the authors suggest may indicate that w-CF outputs are unrealistic.The authors also caution that plausibility needs more robust evaluation.Source: Results and discussion · p. 11

5.2 Experiment 2: Exploring Plausibility and Diversity

• Native Guide counterfactuals were consistently more plausible than w-CF counterfactuals across the evaluated OOD metrics.Table 2 compares the methods using four OOD metrics, with lower percentages of out-of-distribution explanations indicating better results.Source: §5.2 · p. 11, Table 2 · p. 11
• Native Guide can generate diverse counterfactual explanations by using alternative counterfactual instances as guides.Figure 8 shows two explanations for the same query based on perturbing different in-sample counterfactual cases.Source: §5.2 · p. 12, Figure 8 · p. 12

• On diverse UCR datasets with a fully convolutional neural network, Native Guide explanations were significantly better than NUN-CF case-base explanations and w-CF constraint-based explanations.Source: §6 · p. 12
• Native Guide uses existing case-base instances to generate counterfactual candidates that are proximate, sparse, plausible, and diverse.Source: §6 · p. 12
• Methods designed for tabular data often failed to produce meaningful explanations in the time series domain.Source: §6 · p. 12
• Native Guide generates new time series data that the authors say holds promise for data augmentation.Source: §6 · p. 12
• The authors identify experiments with human users and case-based reasoning solutions as future work.Source: §6 · p. 12