Explaining a time‑series model requires both locating the evidence and extracting meaningful knowledge from it. A single timestep is not a semantic unit; the true patterns span many samples, making per‑timestep saliency noisy, high‑dimensional, and unstable. Existing approaches treat fixed‑size windows as independent players in a cooperative game and must define a “baseline” region for a coalition, often using flat lines, noise, or spectral surrogates as fixed counterfactuals. Such choices tend to dominate the explanation and hide the model’s actual decisions.
GRAIL adopts domain‑specific intervals (e.g., the P‑wave and QRS complex of an ECG beat) as players and describes each region with human‑readable features. Counterfactual baselines are generated by an on‑manifold conditional generator, ensuring they lie on the data distribution. Static covariates (age, sex, etc.) enter as singleton players and are weighed directly against temporal regions on the same scale. Predictions are attributed to features via Owen values with common‑random‑number variance reduction.
Experiments on synthetic data and real ECG recordings demonstrate that GRAIL yields faithful and robust explanations under matched on‑ and off‑manifold operators, providing domain experts with actionable knowledge.
Review