{"id":"28664713-dbdb-463b-8dfb-a25329c56d34","arxiv_id":"2505.08748","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Implet extracts contiguous high-attribution subsequences from time series classifiers and clusters them into concise cohort-level explanations.","lead":"This paper introduces Implet, a method that finds short, important segments in time series data by looking at which time steps a model cares about most. It then groups similar segments into clusters to give simpler, higher-level explanations of model decisions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation compares Implets against random subsequences, not against alternative extractors built from the same attributions; the reported faithfulness may be inherited from the attribution step rather than from Implet's subsequence construction.","rationale":"The reader's verdict is CONDITIONAL, and my concern is consistent with that: the paper's central empirical claim needs additional controls before it can be accepted. The reader identified the attribution-accuracy assumption as the weakest link; I agree that this assumption is load-bearing, but my concern is more specific and, in some ways, more fundamental. Even if the attribution scores are perfect, the current evaluation does not show that Implet's particular way of turning attributions into subsequences is what makes the explanations faithful. The comparison against random subsequences merely shows that high-attribution regions matter more than random regions, which is almost a restatement of the attribution assumption. To support the novelty of Implet, the paper must compare against alternative subsequence extraction rules using the same attribution scores. Without that control, the method could be reduced to a trivial threshold scan over any reasonable attribution map, and the claimed contribution of contiguous, length-regularized subsequence extraction would not be validated. The Appendix B protocol change is a secondary but relevant issue: it means the positive results on the hardest datasets were obtained by a different, post-hoc chosen protocol, further reducing confidence in the robustness of the reported effects. These are addressable with additional experiments, so the CONDITIONAL verdict remains appropriate; I would not reject or accept without revision.","tokens_in":12314,"tokens_out":3986,"duration_ms":43607,"concrete_test":"Re-run the Section IV-B faithfulness protocol on the same 13 datasets and two models, adding two baselines built from the same attribution maps: (i) for each sample, remove the fixed-length window of length equal to that sample's median implet length, centered on the timestep with maximum |w_i|; and (ii) remove the non-contiguous set of the |implet| timesteps with largest |w_i|, using the same smooth polynomial replacement. Report paired accuracy drops with error bars across at least 5 random seeds. If either baseline matches or exceeds Implet's mean accuracy drop on a majority of dataset-model combinations, the faithfulness result is attributable to the attribution method, not to Implet's subsequence extraction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section IV-B is that removing implets yields larger accuracy drops than removing random subsequences of equal length. This comparison is too weak to support the paper's claim that Implet 'identifies critical temporal segments' as a novel subsequence explainer. By construction (Section III-A, Eq. 2), an implet is a contiguous span of high cumulative absolute attribution with a length bonus. The ablation removes exactly that high-attribution span, while the random baseline removes an arbitrary span. The observed gap therefore largely revalidates the underlying attribution method rather than testing whether Implet's specific choices—contiguity, length bounds, the lambda length bonus, and the greedy endpoint selection in Algorithm 1—add explanatory value. Section IV-B also states that only one segment is removed per perturbed sample, making the test even less sensitive to the quality of the full extracted set. A baseline that extracts a fixed-length window around the maximum attribution, or a non-contiguous top-k set of the same cardinality, would likely show comparable or larger accuracy drops, because both are driven by the same attribution scores. If that is the case, the distinct contribution of Implet as a subsequence explainer is not established by the provided evidence. The post-hoc protocol change in Appendix B (mean-fill on Chinatown; removing all implets on Earthquakes/FordA) further weakens confidence because the favorable results on initially underperforming datasets are obtained under a different evaluation rule selected after observing the first results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Implet, a post-hoc subsequence explainer for time series classifiers. Implet first computes feature attributions for a target model and then extracts contiguous subsequences whose cumulative absolute attribution plus a length bonus exceeds a threshold, subject to length bounds. A second contribution, Coh-Implet, clusters the extracted subsequences using a two-dimensional dependent DTW distance with DBA centroids and silhouette-based selection of the number of clusters, yielding cohort-level explanations. The method is evaluated on 13 UCR binary classification datasets with FCN and InceptionTime classifiers and six attribution methods. Faithfulness is measured by removing the identified subsequences, using a smooth polynomial replacement scheme, and comparing the accuracy drop with that from removing random subsequences of equal length. The main quantitative claims are that removing implets almost always causes significantly larger accuracy drops than random removal, and that cluster centroids (Coh-Implets) remain faithful when matched by shape alone.","tokens_in":12591,"tokens_out":3286,"duration_ms":37240,"significance":"If the claims are substantiated, Implet provides a simple and inexpensive way to turn arbitrary feature attribution maps into subsequence-level explanations, and the cohort extension is a natural step toward more concise explanations for time series. The paper is clearly written, the algorithms are transparent, and the authors release code. The cohort experiment in Section IV-C is a particularly good idea: matching cohort centroids using only the feature dimension and then measuring faithfulness provides a non-circular check that is stronger than the main ablation. However, the central faithfulness claim currently rests on a comparison against random subsequences, which is too weak to isolate the value of Implet's construction, and the empirical results are reported as point estimates without variance or statistical tests. The manuscript therefore needs additional baselines and a more rigorous experimental protocol before the main claims can be accepted.","major_comments":[{"comment":"The main faithfulness experiment compares removing implets with removing random subsequences of equal length. Because Eq. (2) defines implets as contiguous runs of high cumulative absolute attribution, the observed gap in accuracy drop is largely a consequence of the attribution scores themselves; the experiment does not test whether Implet's specific aggregation choices—contiguity, length bounds, the lambda length bonus, and the greedy endpoint selection in Algorithm 1—add explanatory value. Please add baselines that consume the same attribution scores without these choices, for example a fixed-length window centered on the maximum attribution, a non-contiguous top-k set of the same cardinality, and random contiguous windows sampled from the same length distribution as the extracted implets. The claim that Implet improves over raw attribution maps depends on showing that these alternatives are worse than Implet.","section":"IV-B, Eq. (2)"},{"comment":"The reported accuracy drops are single point estimates with no error bars, confidence intervals, or statistical tests. The claim that removing implets leads to 'significantly larger' accuracy drops is therefore not supported by the presented evidence. Please report repeated runs (for example, multiple model training seeds and multiple random removal seeds), show the variance across samples or runs, and apply a paired test between removal conditions. This is especially important because Fig. 5 shows several datasets where the random and implet removals appear very close (e.g., GunPoint with InceptionTime) and concluding 'almost always' requires knowing how stable the point estimates are.","section":"IV-B, Figs. 5-6"},{"comment":"The evaluation protocol changes for Chinatown, Earthquakes, and FordA are introduced only after the main results in Fig. 5 are found to be unfavorable for those datasets: Chinatown switches to mean-fill removal, and Earthquakes/FordA switch to removing all implets rather than one per sample. This is a post-hoc change in the experimental design, and it is used to recover the desired conclusion. Please either pre-specify the criteria under which the alternative protocol applies, or report both the main and alternative protocols for all datasets so that the comparison remains apples-to-apples. At minimum, the paper should state explicitly that the more favorable results in Fig. 8 are obtained under a different protocol and are not directly comparable to the results for the other ten datasets.","section":"Appendix B"}],"minor_comments":[{"comment":"The abbreviation for the cohort-matched subsequences is inconsistently given as both ICLS (in the text of Section IV-C) and CILS (in the caption of Fig. 6); please unify the notation.","section":"IV-C and Fig. 6"},{"comment":"Line 4 of Algorithm 1 checks 'w_i >= phi' but Eq. (2) accumulates absolute values |w_i|; please clarify whether the threshold in the algorithm is applied to |w_i| or to the raw attribution, and adjust the pseudocode accordingly.","section":"Algorithm 1"},{"comment":"There are several typos and grammatical slips, including 'Shaplets' and 'shaplets' in Sections II and IV-B, 'Earthquaks' in the captions of Figs. 5 and 6, 'subseuqnces' in Appendix B, and 'The last two dataset differ' in Section IV-B; please proofread.","section":"Throughout"},{"comment":"The sentence 'This might due to GuidedBackprop zero-ing out negative gradients' is incomplete; please rephrase.","section":"IV-B"},{"comment":"The paper claims that LASTS is the only prior post-hoc subsequence explainer; given the breadth of the time-series XAI literature, please either provide a more thorough search statement or soften the uniqueness claim.","section":"II-A"}],"recommendation":"major_revision","confidential_remarks":"The central idea is reasonable, but the paper currently overclaims on the basis of a weak baseline and under-powered statistics. The cohort experiment is the most convincing part of the paper; I would encourage the authors to build the evaluation around that and to add attribution-aware baselines in the revision. I do not see a fundamental correctness error that would require rejection, but the current experimental design cannot support the 'significantly larger' claim as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: Implet is a clean, sensible method paper. The combination of feature attribution, contiguous high-attribution subsequence extraction, and DTW/DBA cohort clustering is not in the prior time-series XAI literature I know. The cohort (Coh-Implet) part is the most original piece. But the faithfulness evaluation is weaker than the prose suggests: the main ablation only compares against random subsequence removal, which mostly revalidates the underlying attribution method rather than testing Implet's specific extraction or clustering choices. That said, the weaknesses are addressable, and the paper deserves a proper review.\n\nWhat's new and good: The definition of an implet as a contiguous span of high cumulative absolute attribution with a length bonus (Eq. 2) is simple, clearly specified, and O(T) to compute. The two-dimensional dependent DTW that uses both raw values and attributions is a thoughtful touch for clustering. The CILS experiment in Sec. IV-C is the best evidence in the paper: it finds subsequences matching the cohort centroid by raw shape alone, with no attribution information, and shows they still cause a comparable accuracy drop when removed. That partly answers the circularity worry, because the shape itself carries model-relevant information. The qualitative examples on GunPoint and Chinatown look plausible, and the code is public.\n\nSoft spots: The central claim in Sec. IV-B is that removing implets drops accuracy more than removing random subsequences of equal length. That is a weak baseline. A fixed-length window around the maximum attribution, or a non-contiguous top-k set with the same cardinality, would likely do comparably because all are driven by the same attribution scores. So the specific value added by contiguity, length bounds, the lambda bonus, and the greedy endpoint selection is not established. The figures are also point estimates with no error bars or significance tests. The post-hoc protocol change in Appendix B is legitimately motivated (short series, event-based and frequency-based data), but it was chosen after seeing which datasets failed, and it should be presented as an exploratory robustness check, not as confirmatory evidence. The paper's own observation that InceptionTime results are weaker, and that Earthquakes/FordA are hard for subsequence explainers, is honest but worth deeper analysis.\n\nOverall, I think this is a useful tool paper. The method is plausible, the cohort framework is a genuine contribution, and limitations are openly discussed. It needs a revision with stronger baselines, uncertainty quantification, and more disciplined reporting of the alternative protocols, but it merits a serious referee.","headline":"A clean, well-written method paper that combines attribution maps, contiguous high-score extraction, and DTW/DBA clustering into a genuinely new combination, but the faithfulness evidence needs stronger baselines and statistics before the strong claims hold.","tokens_in":13128,"tokens_out":2750,"would_cite":false,"duration_ms":27599,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Implet finds the exact time segments that drive model predictions.","keywords":["time series explainability","post-hoc explanation","subsequence explanation","feature attribution","cohort explanation","shapelets","faithfulness evaluation","time series classification"],"falsifier":"Take a synthetic time series with a known single decision-relevant motif (e.g., a handcrafted 20-step pattern embedded in noise that fully determines the class label). Train a small CNN, compute Implet with Saliency, and check whether the extracted implet matches the true motif and whether removing it drops accuracy well below random removal; if Implet systematically misses the motif or its removal drop ties random removal, the attribution-to-subsequence assumption is refuted.","tokens_in":12129,"feed_emoji":"📈","tokens_out":7698,"duration_ms":68719,"temperature":0.7,"pith_summary":"The paper proposes Implet, a post-hoc explainer that condenses any per-timestep attribution map for a time series classifier into a small set of contiguous subsequences, called implets. The central claim is faithfulness: in ablation experiments on 13 benchmark datasets, removing the implets a model relies on causes substantially larger accuracy drops than removing random subsequences of equal length. Implet's score for a candidate subsequence is the sum of absolute attribution values plus a length bonus, $\\sum_{i=l}^{r}|w_i| + \\lambda(r-l+1)$, and extraction runs in linear time via a greedy scan. The paper also introduces Coh-Implet, which clusters implets into representative centroids using dynamic time warping, so users can see group-level explanations that remain nearly as faithful as instance-level ones. The motivation is to give deep learning time series classifiers the intuitive, shapelet-like explanations they currently lack, without sacrificing the classifiers' accuracy.","feed_headline":"Implet finds the exact time segments that drive model predictions","feed_subtitle":"Removing the segments it flags cuts accuracy far more than random removal across 13 benchmarks.","key_machinery":"The central object is the implet, a subsequence $I(l,r;x,w)$ paired with its attribution values, scored by $s(l,r)= \\sum_{i=l}^{r} |w_i| + \\lambda(r-l+1)$ and constrained to length between $\\ell_{\\min}$ and $\\ell_{\\max}$. The score's length bonus prevents a single high-attribution spike from being reported, while the cumulative absolute attribution ensures the segment is where the model looks. Extraction is a greedy $O(T)$ scan that finds the best end location for each start position and then skips past the chosen implet. For cohort explanations, the machinery is two-dimensional dependent Dynamic Time Warping (DTW) distance—using both feature values and attribution values—together with DTW Barycenter Averaging (DBA) to compute centroids, and the silhouette score to pick the number of clusters. The removal-based faithfulness evaluation uses a randomized polynomial interpolation to erase a subsequence smoothly, avoiding artificial discontinuities that would confound the ablation.","core_discovery":"On the paper's own terms, Implet establishes that subsequence-level explanations can be both concise and faithful for state-of-the-art time series models. Given a classifier and one input, Implet takes class-specific attribution scores (from any method—saliency, Input×Gradient, DeepLIFT, LIME, SHAP, occlusion) and extracts non-overlapping subsequences whose cumulative absolute attribution plus a length term exceeds a threshold. In the faithfulness test, removing each identified implet with a smooth polynomial interpolation drops accuracy far more than removing random segments in nearly all dataset/model/attribution combinations, and clearly outperforms the ShapeletTransform baseline. Coh-Implet goes further: clustering implets by two-dimensional dependent dynamic time warping and reporting the DTW barycenter as a centroid yields explanations that, when matched to new samples by shape alone, still produce near-implet-level accuracy drops—showing the centroids capture model-relevant structure rather than dataset statistics.","pith_inferences":["I would expect the faithful-subsequence idea to transfer to forecasting and anomaly detection, where identifying the exact temporal window driving a prediction or alert is often more actionable than a per-timestep score; the paper does not test these settings.","The two-dimensional DTW clustering could be repurposed as a post-hoc shapelet discovery tool: the cohort centroids are, in effect, shapelets extracted from a trained model rather than from data statistics, which might give the best of both paradigms.","The smooth polynomial removal procedure is itself a reusable ablation component; future benchmarks could adopt it to reduce artifacts when testing any perturbation-based explainer on smooth or short time series.","A testable extension of the faithfulness claim: on longer multivariate series, generalizing implets to $2n$ dimensions (value plus attribution per channel) may yield much sparser and more precise explanations than per-channel attribution maps alone."],"forward_implications":["Any black-box time series classifier—CNN, InceptionTime, or other—can now be given concise, faithful explanations: a handful of highlighted segments per class instead of a dense attribution vector.","Removal-ablation can become a standard, architecture-aware faithfulness test for time series explainers, and Implet passes it across most UCR benchmarks.","Coh-Implet provides group-level summaries that retain most of the faithfulness of instance-level implets, making it feasible to summarize model behavior over entire datasets or classes.","The method is attribution-agnostic, so it can be layered onto any existing saliency or perturbation explainer to make its output more human-readable.","On frequency-based (FordA) and event-based (Earthquakes) data, subsequence explanations run into a hard limit; the paper maps out where this explanation paradigm should and should not be expected to work."],"supporting_citations":[{"why":"Supplies the 13 benchmark datasets used in all faithfulness evaluations.","marker":"[46]"},{"why":"The InceptionTime architecture whose predictions Implet must explain.","marker":"[1]"},{"why":"The FCN architecture used as the second target model in experiments.","marker":"[33]"},{"why":"The ShapeletTransform baseline that Implet is compared against in removal ablations.","marker":"[19]"},{"why":"Saliency, one of the two attribution backends that give Implet its best faithfulness.","marker":"[35]"},{"why":"Input×Gradient, the other top-performing attribution backend in the faithfulness study.","marker":"[4]"},{"why":"LIME, a perturbation attribution whose sparsity shows why fragmented attributions hurt subsequence extraction.","marker":"[7]"},{"why":"Multidimensional dependent DTW, the distance metric used to cluster implets into cohorts.","marker":"[31]"},{"why":"DTW Barycenter Averaging, the method used to compute cohort centroids.","marker":"[32]"}],"fun_headline_variants":["Implet: expose the time segments that shape predictions","Explain time series models with concise subsequences","Find the exact subsequences that drive model decisions","Uncover critical temporal segments in any time series model","Post-hoc subsequence explainer for time series predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Implet assumes that feature attribution scores accurately reflect which time steps drive the model's prediction, so that a contiguous stretch of high absolute attribution marks a genuinely decision-relevant subsequence.","fun_headline_variants_meta":{"raw":{"variants":["Implet: expose the time segments that shape predictions","Explain time series models with concise subsequences","Find the exact subsequences that drive model decisions","Uncover critical temporal segments in any time series model","Post-hoc subsequence explainer for time series predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000316,"raw_usage":{"total_tokens":1741,"prompt_tokens":847,"completion_tokens":894,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":821}},"tokens_in":463,"tokens_out":894,"duration_ms":8468,"temperature":1.0,"reasoning_tokens":821,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:46:51.077798+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic time series with a known single decision-relevant motif (e.g., a handcrafted 20-step pattern embedded in noise that fully determines the class label). Train a small CNN, compute Implet with Saliency, and check whether the extracted implet matches the true motif and whether removing it drops accuracy well below random removal; if Implet systematically misses the motif or its removal drop ties random removal, the attribution-to-subsequence assumption is refuted.","supporting_citations":[{"cited_title":"Lstm fully convolu- tional networks for time series classification,","cited_arxiv_id":null,"evidence_quote":"The FCN architecture used as the second target model in experiments."},{"cited_title":"A shapelet transform for time series classification,","cited_arxiv_id":null,"evidence_quote":"The ShapeletTransform baseline that Implet is compared against in removal ablations."},{"cited_title":"“why should i trust you?","cited_arxiv_id":null,"evidence_quote":"LIME, a perturbation attribution whose sparsity shows why fragmented attributions hurt subsequence extraction."},{"cited_title":"Generalizing dtw to the multi-dimensional case requires an adaptive approach,","cited_arxiv_id":null,"evidence_quote":"Multidimensional dependent DTW, the distance metric used to cluster implets into cohorts."},{"cited_title":"A global averaging method for dynamic time warping, with applications to clustering,","cited_arxiv_id":null,"evidence_quote":"DTW Barycenter Averaging, the method used to compute cohort centroids."}],"review_version":1}