A preprint reports that its time-series explanation method beat competing attribution explainers in nine of 12 synthetic evaluation cases, while its counterfactual mode was reported to achieve the highest validity and confidence across four untargeted datasets. In plain language, feature attribution highlights important parts of a sequence, while a counterfactual explanation generates an altered sequence aimed at a different model output. The system is called TimeX++, with TimeXa++ for attribution and TimeXcf++ for counterfactuals.
Both modes minimize the same overarching objective through end-to-end optimization. Its architecture combines a shared attribution bottleneck extractor with separate conditioned generators for the attribution and counterfactual modes. Put simply, the paper treats the Information Bottleneck as a shared training rule for deciding what information an explanation should retain.
Attribution scores favored the new method
On the synthetic tests, the paper reports average gains over TimeX of 11.01% in AUPRC, 10.87% in AUP and 1.25% in AUR, scores that compare the highlighted features with the salient features built into the synthetic tests. The reported Friedman statistic was 51.32, with p < 0.001.
Real-world benchmark tests used occlusion, an indirect check in which parts of the input are hidden or replaced before the classifier is evaluated. The proposed method was reported to outperform existing explainers under different perturbations, and zero substitution produced the highest average rank. The reported Friedman statistic was 25.80, with p < 0.001.
Counterfactuals balanced several demands
Counterfactual tests considered four competing qualities: validity, confidence, sparsity and proximity. In targeted evaluation, TimeXcf++ was reported to balance validity and confidence against sparsity and proximity. On the Epilepsy dataset, it had the highest confidence with average sparsity of 0.14.
In untargeted evaluation, TimeXcf++ was reported to achieve the highest validity and confidence across all four datasets.
The tests also examined distribution and noise
Another set of tests examined whether explanations remained aligned with the original data distribution. TimeXa++ explanations were reported to have lower MMD and KL divergence than conventional masking strategies, two measures used to compare distributions. That result points to closer distributional alignment on the reported tests.
At a 50% noise ratio, TimeXcf++ retained validity above 0.8 on both tested datasets, according to the paper's reported results. Validity and confidence decreased as the injected noise increased.
Efficiency and tuning revealed trade-offs
The reported efficiency results separated training from inference. Training was reported to take approximately 3 to 15 minutes per fold, while generating counterfactuals at inference was reported to be over 100 times faster than the fastest counterfactual baseline. The training figures were described as approximate.
The paper's tuning analysis exposed a trade-off. Attribution performance was reported to stabilize when the hyperparameter r was between 0.4 and 0.7. For counterfactuals, increasing r was reported to improve validity and confidence but worsen sparsity and proximity.
Ablation tests that removed parts of the objective also changed the reported counterfactual profile. Removing the reference-distance or structural-anchor losses was reported to worsen proximity and sparsity, while the distribution-maintenance loss was reported to have a smaller but consistent distributional effect.
The comparison covered several models
The study benchmarked TimeX++ against established attribution and counterfactual explainers from both paradigms.
The reported pattern also held across classifier architectures. With LSTM and CNN predictors, TimeXa++ retained the best reported overall attribution performance, while TimeXcf++ maintained a balance across counterfactual metrics.
The evaluation used ten synthetic and real-world time-series benchmark datasets. The document is identified as arXiv:2608.25897v1, dated 26 Aug 2026, and labeled a preprint. The evidence reported here is therefore the performance measured in those benchmark tests.
Paper data and sources
Original title: Towards A Unified Information Bottleneck Framework for Time Series Explanations
Authors: Xu Zheng, Zichuan Liu, Zhuomin Chen et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text