{"id":"fab901f2-3ce5-438e-81ea-9d3433a74a52","arxiv_id":"2509.03649","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Equal-length segmentation matches or beats seven more complex segmentation strategies for SHAP-based time series explanations, and length-weighted normalisation of segment attributions improves evaluation scores.","lead":"This paper compares eight ways of splitting time series into segments before computing SHAP explanations for classifiers, and finds simple equal-length splits work as well or better than more complex segmentation algorithms. It also proposes dividing each segment's attribution among its time points, a normalisation step that improves scores on two XAI evaluation benchmarks in most settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that segment count matters more than segmentation method is not tested: no experiment varies segment count independently of method.","rationale":"The reader's weakest assumption concerned the validity of InterpretTime and AUCD, particularly for non-deep-learning classifiers. That is a serious concern, acknowledged by the authors themselves in Section 5.3. However, I identify a different, more fundamental gap: the paper's central novel claim about segment count versus method is not directly tested because segment count is never isolated as a factor. The reader's rationale did mention this issue, but selected the metric fragility as the weakest assumption. I argue the confound is more load-bearing because even if the metrics were perfect, the headline claim would still lack support. My concern does not change the overall verdict—the paper remains a useful empirical study with plausible practical recommendations, and the released code could enable the requested re-analysis. The concrete test would either validate the claim or require the authors to soften the abstract. I agree with the reader that the paper is CONDITIONAL, not fundamentally flawed, so the verdict is unchanged.","tokens_in":9644,"tokens_out":4541,"duration_ms":47143,"concrete_test":"Run a controlled study on one dataset/classifier pair (e.g., MiniRocket/Gunpoint as in Fig. 4a). Fix the equal-length method and compute InterpretTime AUCSE and AUCD for n = 2, 4, 8, 16, 32 segments, holding background and perturbation fixed. Then compute the variance in scores across n, and compare it to the variance across the eight segmentation methods when all are forced to the same segment count (e.g., n = 10). If the across-n variance is not larger than the across-method variance at fixed n, the headline claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Section 6 assert that 'the number of segments has a greater impact on explanation quality than the specific segmentation method' and that equal-length segmentation is consistently better. However, Section 5's conditional variance analysis treats only three factors—background, perturbation, and segmentation method—and never manipulates the number of segments as a controlled variable. Section 4.2 lists hyperparameters for the eight methods but gives no indication that segment counts were matched across methods or varied within a method. If methods produce different numbers of segments (ClaSP and NN-Segment detect change points; equal-length uses an arbitrary n), then observed 'method' differences are confounded with segment-count differences, and the headline conclusion 'segment count matters more' is not identifiable from the described experiments. This is an internal-validity gap in the paper's central novel claim, independent of the acknowledged fragility of the evaluation metrics for non-deep classifiers. The equal-length 'winner' claim is also method-dependent: if n is chosen arbitrarily for equal-length while other methods auto-select their own counts, the comparison may be unfair.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates how the choice of time-series segmentation affects the quality of SHAP-based explanations for time-series classifiers. It compares eight segmentation methods (equal-length, ClaSP, Information Gain, Greedy Gaussian, NNSegment, Binary Segmentation, Bottom-up, KernelCPD) on four classifiers and five datasets, with two SHAP background strategies and two XAI evaluation protocols (InterpretTime and AUCD). The abstract and conclusion make two headline claims: (1) the number of segments matters more than the specific segmentation method, and (2) equal-length segmentation consistently outperforms most custom segmentation algorithms. The paper also proposes a length-based normalisation of segment-level SHAP values to preserve segment-level additivity, and reports that this normalisation consistently improves attribution-quality scores.","tokens_in":9906,"tokens_out":3040,"duration_ms":34924,"significance":"If the claims are correct, the paper would provide actionable practical guidance for the growing area of segment-based SHAP for time series: users could adopt cheap equal-length segmentation and a simple normalisation rule to obtain equal or better explanations than with sophisticated segmentation algorithms. The paper is well-positioned in the literature and makes its code and data publicly available, which supports reproducibility. However, the headline contribution—the relative importance of segment count versus segmentation method—is not operationalised in the experimental design, and the evaluation metrics are acknowledged by the authors themselves to be unreliable for three of the four classifiers. The proposed normalisation is justified by a definitional additivity argument, but its empirical benefit is measured with the same fragile metrics. These issues are central to the paper's claims, so the work currently requires substantive revision.","major_comments":[{"comment":"The central claim that 'the number of segments has a greater impact on explanation quality than the specific segmentation method' is not testable from the described experiments. Section 5's conditional variance analysis considers backgrounds, perturbations, and segmentation methods, but never varies the number of segments independently of the method. Section 4.2 lists hyperparameters for each segmentation, yet gives no indication that segment counts were matched across methods or varied within a method. Because each method produces its own segment count (e.g., ClaSP detects change points, equal-length uses an externally set n), the observed differences between 'methods' inevitably conflate method identity with segment count. The conclusion in Section 6 therefore goes beyond what the data can identify.","section":"Abstract, Section 5, Section 6"},{"comment":"The justification for the proposed normalisation is not sufficient. Section 3 asserts that preserving segment-level additivity is 'essential' and that the proposed uniform distribution 'ensures that evaluations performed on individual timepoints remain faithful to the original segment-wise Shapley values.' This is a definitional property, not an empirical or theoretical argument that additivity at the segment level is the correct criterion for timepoint-level explanation quality. The empirical claim that normalisation 'consistently improves attribution quality' is based on InterpretTime and AUCD scores, which rank individual timepoints; changing the scale by dividing by segment length directly changes cross-segment rankings for uneven segment lengths. The paper does not rule out that the improvement is an artefact of the metric rather than a genuinely better explanation. At minimum, the","section":"Section 3, Section 5.2"},{"comment":"The paper's own limitation statement undermines the generality of the empirical results. It reports that the noise-based training augmentation required by InterpretTime degrades non-deep-learning classifiers, and that without it, results 'may vary unpredictably across different perturbations.' Since QUANT, MiniRocket, and RandomForest are used in a large part of the experimental matrix, the ranking-based conclusions for these classifiers rest on a fragile evaluation foundation. The paper concludes that InterpretTime's effectiveness 'may be limited to deep learning classifiers,' yet the headline recommendations about segmentation and normalisation are made across all classifiers rather than only for ResNet. This inconsistency needs to be resolved, for instance by restricting the claims to deep classifiers or by validating the metrics further.","section":"Section 5.3, Section 6"},{"comment":"The claim that 'equal-length segmentation consistently outperforms most of the custom time series segmentation algorithms' is not supported by any direct statistical comparison. The equal-length segmentation's number of segments n is not specified in Section 4.2, so it is unclear whether the comparison is fair: a poorly chosen n could handicap equal-length, while the custom methods autodetect their segment counts. Section 5's variance analysis is not a substitute for a per-dataset, per-classifier pairwise comparison with a statistical test. Figure 1 shows error bars but no significance testing is reported anywhere in Section 5. The 'consistently' qualifier is too strong based on the evidence presented.","section":"Section 4.2, Section 5"}],"minor_comments":[{"comment":"Typo: 'Metholodologies' should be 'Methodologies'.","section":"Section 4.3"},{"comment":"The spelling alternates between 'normalisation' (Section 3) and 'normalization' (Section 5.2 and elsewhere). Please standardise.","section":"Section 3 / 5.2"},{"comment":"The description of NNSegment says it 'comes from LimeSegment' but the original method is cited as [26]; consider stating the relationship more explicitly.","section":"Section 2.2"},{"comment":"The phrase 'equal-length segmentation is often the best choice' is justified by computational cost rather than by explanation quality; this is an argument about efficiency, not about the quality claim, and should be separated.","section":"Section 5"},{"comment":"The caption mentions 'normalised entropy' but the body text uses 'relative entropy'; please align the terminology.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practical question, and the code release is a strength. However, the two headline claims—segment count dominance and the superiority of equal-length segmentation—are not properly isolated in the experimental design, and the evaluation metrics are acknowledged to be unreliable for a large subset of the classifiers. The most straightforward remedy is to add experiments that vary the number of segments within a fixed method (e.g., equal-length with different n values, or a change-point method with a user-specified number of change points), and to report results separately for deep and non-deep classifiers. If such experiments are infeasible, the claims should be substantially weakened. The normalisation contribution is interesting but needs a stronger theoretical or empirical justification that it improves explanation quality rather than merely shifting metric scores."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2509.03649. The genuinely useful part is the systematic comparison of eight segmentation strategies for segment-based SHAP on time series classification. Prior work mostly used equal-length segmentation; this paper gives a reasonably broad experimental matrix (five datasets, four classifiers, two backgrounds, two evaluation metrics) and shows that equal-length segmentation is usually as good as or better than bespoke change-point methods, at much lower cost. The length-based normalisation is a simple, sensible fix: distributing each segment's Shapley value uniformly over its time points preserves segment-level additivity. The result that this rescaling helps most when segment lengths are uneven is plausible and empirically demonstrated, especially on the EOG and KLC datasets. Credit where due: the code and data are released, the comparison is broad, and the practical guidance is actionable.\n\nThe soft spots are real, though. The headline claim that 'the number of segments has a greater impact than the segmentation method' is not actually supported by the experiments as described. No experiment varies segment count independently of method. Different methods produce different numbers of segments (ClaSP and NNSegment auto-detect change points; equal-length uses an arbitrary n), so observed method differences are confounded with segment-count differences. This is an internal-validity gap in the central claim. As far as I can see, the paper only shows that 'segmentation method' explains little variance overall, which is not the same thing. A controlled study where each method is run with matched segment counts (or where segment count is swept within a fixed method) would be needed.\n\nSecond, the evaluation metrics for non-deep-learning classifiers are fragile. InterpretTime requires noise-based training augmentation in its original form; the authors themselves state they omitted it because it degrades accuracy for QUANT, MiniRocket, and RandomForest, and note results 'may vary unpredictably' across perturbations. Since those classifiers are a large part of the experimental matrix, the ranking-based conclusions rest on a foundation the paper itself acknowledges is shaky. AUCD is more robust in this respect, but it's a different proxy, and the two metrics sometimes disagree.\n\nThe normalisation contribution is more solid: it's derived from an additivity property, tested externally via the two metrics, and shows a clear pattern. The background comparison is minor but fine.\n\nOverall, this is a useful benchmark paper for the time series XAI subfield. It deserves a serious referee, but the referee should push for a direct segment-count experiment and a more careful handling of the evaluation metrics for non-DL classifiers. The equal-length recommendation may hold up, but the 'number of segments matters more' claim needs proper support.","headline":"A useful empirical benchmark on segmentation choices for SHAP on time series, but the central 'segment count matters more than method' claim is not actually tested as stated.","tokens_in":10333,"tokens_out":662,"would_cite":true,"duration_ms":7945,"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":"For SHAP explanations of time series, the simple equal-length segmentation beats most custom segmentation algorithms.","keywords":["time series classification","SHAP","segmentation","explainable AI","attribution normalization","InterpretTime","AUCD","XAI evaluation"],"falsifier":"Repeat the segmentation comparison on the same datasets and classifiers using a fidelity measure that does not depend on noise-augmented retraining, for example deletion/insertion with the original model across multiple random seeds, or a domain-expert oracle on Gunpoint and KeplerLightCurves. If equal-length segmentation no longer ranks in the top group under that metric, the paper's central conclusion does not transfer.","tokens_in":9568,"feed_emoji":"📊","tokens_out":5280,"duration_ms":52424,"temperature":0.7,"pith_summary":"The paper tries to show that when SHAP is applied to time series through segment-based feature aggregation, the number of segments matters more than the segmentation algorithm that produces them. It compares eight segmentation methods across five datasets and four classifiers and finds that equal-length segmentation scores as well as or better than most custom change-point-detection approaches, at a fraction of the cost. The paper also argues that dividing each segment's SHAP value by the segment length, preserving the Shapley additivity property at the timepoint level, consistently improves attribution scores under both evaluation measures. A reader should care because the paper offers a cheaper and more defensible default for explaining time series classifiers, while flagging that the evaluation methods themselves are not neutral.","feed_headline":"Equal-length segmentation wins for time-series SHAP","feed_subtitle":"Number of segments matters more than the algorithm; a length-weighting fix improves attributions further.","key_machinery":"The load-bearing objects are: (1) segment-based SHAP, where consecutive timepoints are grouped into features so the Shapley computation becomes tractable; (2) the eight segmenters, with equal-length segmentation as the cost-free baseline; (3) the length-weighted normalisation phi_S/|S|, which restores segment-level additivity after a segment's value is spread over its timepoints; and (4) the two evaluators, InterpretTime (AUCSE and F-score) and AUCD, which perturb time series according to attribution rankings to measure explanation quality.","core_discovery":"The central discovery is that segment-level SHAP explanations for time series do not need elaborate segmentation algorithms: the number of segments is the dominant factor in explanation quality, and equal-length segmentation is the best default, matched by most custom methods only when they produce similar partitions. A second, more general finding is that attributions remain theoretically sound only if segment-level Shapley values are distributed uniformly over the timepoints of each segment, dividing the segment's value by its length; doing so preserves segment-level additivity and empirically raises explainability scores, with the largest gains on the most uneven segmentations, such as th","pith_inferences":["If the ranking holds under other fidelity metrics, the whole category of 'smart' segmentation for SHAP may be over-engineered; future work could redirect effort from change-point detectors to choosing segment counts and validating background sets.","The length-normalisation insight generalises beyond SHAP: any group-based attribution method that assigns the same weight to every member of a group should scale by group cardinality to preserve additivity, which is a testable extension for model-agnostic explainers on tabular data.","The observed disagreement between InterpretTime and AUCD for non-deep classifiers implies that comparative XAI papers in time series should report both metrics or re-validate with an evaluation method that does not retrain with noise augmentation.","A cheaper practical default could be quantified: the paper's data suggests a plateau in explanation quality at a relatively small number of equal-length segments, but it does not derive a formula for the minimal segment count."],"forward_implications":["Practitioners using segment-based SHAP should default to equal-length segmentation; fancier change-point detectors only help when they produce materially different partitions, and they come at higher computational cost.","The length-weighting normalisation should become standard for segment-based explanations: it improves evaluation scores, especially for uneven segmentations, and leaves equal-length explanations essentially unchanged.","The zero background is an acceptable default, but when it differs from the average background, the average background typically gives better or equal explanation quality.","The choice of evaluation measure matters: InterpretTime and AUCD can disagree, particularly for non-deep-learning classifiers, so conclusions about explanation quality should be qualified by the evaluation method used.","The number of segments, not the choice of segmenter, is the primary lever on explanation quality; practitioners should set the segment count first and treat algorithm choice as secondary."],"supporting_citations":[{"why":"Establishes SHAP as the attribution method being applied to segmented time series.","marker":"[17]"},{"why":"Supplies the segment-based Shapley value sampling implementation, the equal-length baseline, and the extended InterpretTime evaluation used in the experiments.","marker":"[25]"},{"why":"Introduces InterpretTime, the primary evaluation methodology, including the noise-augmented training that the paper finds problematic for non-deep classifiers.","marker":"[28]"},{"why":"Provides the deletion/insertion protocol and the AUCD metric used as the second evaluation measure.","marker":"[19,21]"},{"why":"Supplies the ClaSP change-point detector, one of the custom segmentation methods compared against equal-length segmentation.","marker":"[8]"},{"why":"Supplies the NNSegment change-point detector adapted from LimeSegment, another custom segmentation method in the comparison.","marker":"[26]"}],"fun_headline_variants":["Segment count drives SHAP quality, not algorithm","Equal-length segmentation wins for SHAP","Length-weighting fixes SHAP attributions","SHAP explainability: simple segmentation beats complex","Segment count matters more than method for SHAP"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The ranking of segmentation methods stands or falls with the validity of the two evaluation scores (InterpretTime and AUCD) as measures of true explanation quality; the paper itself observes that InterpretTime's required noise augmentation degrades non-deep-learning models and that its results can vary unpredictably across perturbations for those models.","fun_headline_variants_meta":{"raw":{"variants":["Segment count drives SHAP quality, not algorithm","Equal-length segmentation wins for SHAP","Length-weighting fixes SHAP attributions","SHAP explainability: simple segmentation beats complex","Segment count matters more than method for SHAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1144,"prompt_tokens":726,"completion_tokens":418,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":350}},"tokens_in":470,"tokens_out":418,"duration_ms":4549,"temperature":1.0,"reasoning_tokens":350,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:46:16.557873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the segmentation comparison on the same datasets and classifiers using a fidelity measure that does not depend on noise-augmented retraining, for example deletion/insertion with the original model across multiple random seeds, or a domain-expert oracle on Gunpoint and KeplerLightCurves. If equal-length segmentation no longer ranks in the top group under that metric, the paper's central conclusion does not transfer.","supporting_citations":[{"cited_title":"Advances in neural information processing systems30 (2017)","cited_arxiv_id":null,"evidence_quote":"Establishes SHAP as the attribution method being applied to segmented time series."},{"cited_title":"In: Machine Learning and Knowledge Discovery in Databases","cited_arxiv_id":null,"evidence_quote":"Supplies the segment-based Shapley value sampling implementation, the equal-length baseline, and the extended InterpretTime evaluation used in the experiments."},{"cited_title":"Nature Machine Intelligence5(3), 250–260 (2023)","cited_arxiv_id":null,"evidence_quote":"Introduces InterpretTime, the primary evaluation methodology, including the noise-augmented training that the paper finds problematic for non-deep classifiers."},{"cited_title":"Data Mining and Knowledge Discovery 37(3), 1262–1300 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the ClaSP change-point detector, one of the custom segmentation methods compared against equal-length segmentation."},{"cited_title":"In: Camps-Valls, G., Ruiz, F.J.R., Valera, I","cited_arxiv_id":null,"evidence_quote":"Supplies the NNSegment change-point detector adapted from LimeSegment, another custom segmentation method in the comparison."}],"review_version":1}