{"id":"6e2655ae-c3f6-42e5-a9bf-0202e02a81eb","arxiv_id":"2505.23624","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EMeriTAte+DF converts multivariate time series into polyadic event logs, mines explainable dataful declarative rules, and uses them to classify time series with strong but not universal benchmark results.","lead":"This paper describes EMeriTAte+DF, a pipeline that turns numerical time series into event-style logs, learns logical rules that explain each class, and classifies new series with those rules. It reports high accuracy on several time-series benchmarks, but it does not consistently beat every state-of-the-art method it compares against.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract's unqualified 'outperform state-of-the-art' claim is contradicted by the paper's own Table 4, where EMeriTAte+DF scores below several competitors on Italy Power Demand and Basic Motions.","rationale":"The reader's weakest_assumption focuses on pattern pruning removing class-discriminative information, which is a plausible mechanism for the Italy Power Demand drop and is explicitly acknowledged in §4.1 and §5.2. However, the more fundamental load-bearing concern is that the abstract's general 'outperform SOTA' claim is already falsified by the paper's own Table 4 on Basic Motions, where pruning does not explain the gap (CIF and TapNet score 100.0% vs. 98.75%). This is an internal-consistency problem, not a disagreement with external consensus, and it directly targets the strongest claim identified by the reader. The methodological contribution—polyadic logs and dataful Poly-DECLARE clause mining—remains plausible and potentially novel, but the empirical evidence as reported does not support the unqualified superiority claim. The appropriate remedy is empirical revision: report significance tests, correct the '≈99%' summary, and replace the blanket claim with dataset-specific statements. Since the required fix is revision of the empirical claims rather than a fundamental flaw in the pipeline, conditional acceptance remains appropriate; our stress test does not move the reader's verdict.","tokens_in":20780,"tokens_out":5281,"duration_ms":50093,"concrete_test":"Run a paired McNemar test (or exact binomial sign test) on the 10 stratified train/test folds for Italy Power Demand, comparing EMeriTAte+DF against Rocket on the same folds. If the difference is not statistically significant at α=0.05, or if Rocket is favored, the abstract's unqualified 'outperform state-of-the-art' claim must be weakened to a dataset-specific or 'competitive' claim; if the difference favors EMeriTAte+DF, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—'Our previous and current solutions outperform state-of-the-art solutions for multivariate time series classifications'—is directly contradicted by the reported results. In Table 4(a) (Italy Power Demand), EMeriTAte+DF achieves 96.23% accuracy versus Rocket 97.02%, CIF 96.69%, STC 96.81%, and the previous EMeriTAte 99.59%. In Table 4(b) (Basic Motions), CIF and TapNet both reach 100.0% while EMeriTAte+DF gets 98.75%. The text in §5.2 further asserts that when EMeriTAte+DF did not achieve maximum scores it still attained ≈99% accuracy, which is inconsistent with the reported 96.23% on Italy Power Demand. No statistical significance tests are provided; the '±' values are max-min ranges over 10 runs rather than standard deviations or confidence intervals, so the magnitude of any advantage cannot be assessed. Because the abstract's superiority claim is unqualified and the paper's own evidence refutes it on two of four datasets, the central claim is unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents EMeriTAte+DF, a pipeline for explainable multivariate time series classification. It discretizes MTS segments into polyadic logs whose durative constituents carry Catch22 payloads, mines a dataful extension of DECLARE clauses over concurrent constituents, uses clause satisfaction/violation/vacuous satisfaction to build an embedding, and trains a decision tree to obtain a white-box classifier. The authors claim that the new solution and its predecessor outperform state-of-the-art MTSC classifiers and that the approach yields human-readable explanations. Evaluation is performed on four datasets: Italy Power Demand, Basic Motions, OSULeaf, and a Parkinson's disease dyskinesia dataset.","tokens_in":21028,"tokens_out":4062,"duration_ms":38608,"significance":"If the accuracy and explainability claims were borne out, this would be a valuable bridge between numerical MTSC and declarative process mining. The paper contains a precise algorithmic framework (polyadic logs, taxonomies, polyadic DECLARE semantics, dataful refinement via decision trees), and it cites a public code release. The main empirical claim, however, is not supported by its own reported numbers, so the current version overstates the contribution.","major_comments":[{"comment":"The abstract's unqualified claim that 'our previous and current solutions outperform state-of-the-art solutions' is contradicted by the paper's own Table 4. On Italy Power Demand (Table 4a), EMeriTAte+DF scores 96.23% accuracy, below Rocket (97.02%), CIF (96.69%), STC (96.81%), and the previous EMeriTAte (99.59%). On Basic Motions (Table 4b), CIF and TapNet reach 100.0% while EMeriTAte+DF reaches 98.75%. The text in §5.2 further states that when EMeriTAte+DF did not achieve maximum scores it still achieved about 99% accuracy, which is inconsistent with the reported 96.23% on Italy Power Demand. The superiority claim must be removed or substantially qualified, and the internal inconsistency resolved.","section":"Abstract and §5.2, Table 4"},{"comment":"No statistical significance tests are provided. The reported '±' values are maximum-minus-minimum ranges over 10 runs, not standard deviations or confidence intervals. Since several differences are small (e.g., 96.23% vs 96.81% on Italy Power Demand), the reader cannot determine whether EMeriTAte+DF is competitive at all. Please report standard deviations or confidence intervals and, ideally, pairwise significance tests (e.g., McNemar or a signed-rank test) against the main competitors.","section":"§5.2, Fig. 4"},{"comment":"The pruning of 'high volatility patterns' appears to be chosen after inspecting results. Section 4.1 states that some high volatility patterns are discarded to favor shorter volatility patterns, and §5.2 attributes the Italy Power Demand shortfall to fluctuations that were not discarded in the previous solution. This makes the claimed superiority difficult to falsify: a different dataset-specific pruning could reverse the comparison. Please either justify the pruning through a validation-based procedure or clearly present it as a dataset-specific limitation, and report an ablation with and without the pruning.","section":"§4.1 and §5.2"},{"comment":"Algorithm 5 uses the notation π(ς) for the span of a constituent in CHAINS (line 4) and in RESPPREC (line 29), but π is never defined in that algorithm. If π is meant to be the span function δ introduced in §3.1, this should be stated explicitly; otherwise the mining procedure is underspecified and cannot be reproduced from the pseudocode.","section":"Algorithm 5"}],"minor_comments":[{"comment":"There is an unresolved placeholder 'Section??' in the paragraph describing the main mining algorithm; this should be replaced with the actual section reference.","section":"§4.1"},{"comment":"The text refers to a 'Catch24-payload' after previously defining Catch22 features; this is presumably a typo and should be corrected.","section":"§4.1"},{"comment":"The proof of Lemma 2 gives the bound 'O(s^2 log s C_s t/s N)', but the meaning of the term t/s is unclear and the claimed 'superpolynomial' worst-case behavior is not reconciled with the polynomial-looking expression for C_s. Please state the assumptions on C_s explicitly and give a cleaner closed-form bound.","section":"§4.1, Lemma 2"},{"comment":"The statement that 'we kept different training/testing splits from the ones in our previous paper' is not accompanied by the actual split specifications; please provide the exact splits or a randomization seed for reproducibility.","section":"§5.2"},{"comment":"The accuracy differences between EMeriTAte+DF and competitors on Italy Power Demand are within roughly one percentage point, so the visual coloring of best/worst results should be supplemented by a discussion of effect size and variability rather than only point estimates.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The core algorithmic contribution is original and the code release is a plus, but the empirical part needs a serious recalibration. The overclaim in the abstract is a load-bearing issue that can be fixed by tempering the claim and adding proper statistical comparisons. I do not see evidence of misconduct; the issue is that the reported numbers do not support the stated conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the paper's real contribution is a detailed extension of declarative process mining to polyadic logs with concurrent, dataful constituents, plus a class-refinement step based on decision trees. That is worth your time. The abstract, however, claims the method outperforms state-of-the-art MTS classifiers, and the paper's own results say otherwise on two of four datasets. The empirical story needs a rewrite before this is publishable.\n\nWhat is genuinely new: EMeriTAte+DF extends the earlier EMeriTAte by adding Catch22 payloads to mined data-trend constituents, mining polyadic DECLARE clauses that allow multiple constituents at the same timestamp, and refining those clauses via decision-tree predicates that separate classes. The algorithms are stated precisely, with complexity lemmas, and the design choices are motivated by shortcomings in DML and Bolt2. That is a real step forward for explainable MTS classification, and I can see follow-up work using the polyadic dataful clause formalism.\n\nWhere it falls down: the abstract's superiority claim is contradicted by Fig. 4. On Italy Power Demand, EMeriTAte+DF gets 96.23% while Rocket gets 97.02% and the earlier EMeriTAte gets 99.59%. On Basic Motions, CIF and TapNet reach 100% and EMeriTAte+DF gets 98.75%. The text in §5.2 even says that when EMeriTAte+DF did not reach maximum scores it still attained ~99% accuracy, which is flatly inconsistent with the Italy Power Demand result. That needs to be fixed and toned down. Second, the reported ± intervals are max-min ranges over ten runs, not standard deviations or confidence intervals; there are no significance tests, so we cannot tell whether any of the differences matter. Third, the Dyskinetic dataset has only two MTS and the clinical data is not public; the 100% result there is suggestive but not a strong benchmark. Fourth, several hyperparameters (pruned data-trend patterns, max tree depth, purity thresholds) appear to have been chosen after seeing results; without a sensitivity analysis, that weakens the evaluation. To be clear, the core pipeline is not circular: discretization is unsupervised, and the refinement uses training labels but is evaluated on held-out splits. The problem is not the logic of the method, it's the strength of the empirical claims.\n\nFor a reader: someone working on declarative process mining or explainable time series classification will find the formalization and the mining algorithms valuable. The empirical claims need major revision, but the underlying idea is sound enough to deserve serious peer review. I would send it to reviewers, with the expectation that the authors will revise the abstract, correct the §5.2 inconsistency, add statistical rigor, and release the clinical data or a detailed description. If they do that, it should become a solid contribution.","headline":"Solid formal extension of DECLARE to polyadic dataful logs, but the abstract's 'outperform SOTA' claim is refuted by the paper's own tables on two of four datasets.","tokens_in":21531,"tokens_out":3710,"would_cite":true,"duration_ms":33881,"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":"By discretizing multivariate time series into polyadic event logs and mining data-enriched Poly-DECLARE temporal clauses, this paper builds a classification pipeline whose decisions are both accurate and human-readable.","keywords":["Verified AI","eXplainable AI (XAI)","polyadic logs","Data Trends","Poly-DECLARE","Multivariate Time Series Classification","specification mining","Catch22"],"falsifier":"Run EMeriTAte+DF on Italy Power Demand with the currently discarded high-volatility patterns included in the DT-mining step: if accuracy returns to the roughly 99.6% level of the dataless EMeriTAte while runtime stays competitive, then the pruning rule, not the dataful refinement, is what limits this dataset, and the general claim that the pipeline retains all class-discriminative information is falsified.","tokens_in":20587,"feed_emoji":"📈","tokens_out":9382,"duration_ms":81847,"temperature":0.7,"pith_summary":"This paper proposes EMeriTAte+DF, a pipeline that rewrites multivariate time series (MTS) into polyadic event logs, where each timestamp can hold several concurrent constituents that carry activity labels, duration spans, and numerical payloads, and then mines 'dataful' Poly-DECLARE temporal-logic clauses from those logs. The central claim is that the mined clauses, scored by whether each trace satisfies, violates, or vacuously satisfies them and then fed to a decision tree, yield an accurate classifier for MTS whose decisions are human-readable rules about temporal correlations. The authors report accuracy in the range of about 96–100% across four datasets and argue that the approach outperforms state-of-the-art MTSC classifiers, with the largest margin on a Parkinson's dyskinesia dataset where competing methods score below 26%. Because the clauses can be refined with numerical predicates learned from payloads, classes that share the same bare temporal skeleton can still be separated, which is the step the paper says makes the explanation genuinely dataful.","feed_headline":"Mining time series into readable rules beats state of the art","feed_subtitle":"Mining multivariate sensor data into declarative rules keeps accuracy high and makes every classification explainable.","key_machinery":"The load-bearing object is the polyadic log model paired with the Poly-DECLARE specification language. A polyadic log is a trace collection where each timestamp can hold several durative constituents simultaneously, each marked with an activity label, a numerical payload, and a time span; this is what lets a multivariate numeric series become an event log without forcing simultaneous phenomena into one label. Poly-DECLARE extends the DECLARE templates (Init, Exists, Precedence, Response, ChainResponse, plus 'All' variants that require all constituents in an event to satisfy the condition) to test constraints on these concurrent constituents and to carry data predicates on activations and targets. The decisive mechanism is the refinement step: when the same dataless clause is mined from two class-segmented logs, a decision tree is trained on the payloads of the activating (or target) constituents, and each tree path becomes a unary predicate that specializes the clause; if no predicate separates the payloads with enough purity, the algorithm backtracks to the dataless clause. This yields class embeddings of clause satisfactions, violations, and vacuous satisfactions, over which a final decision tree produces the human-readable propositional class description.","core_discovery":"On the paper's own terms, the central discovery is that the separation between numerical time series analysis and event-based declarative process mining can be closed: every multivariate time series segment can be described as a polyadic trace, and Poly-DECLARE clauses mined from these traces double as both classifiers and explanations. The dataful extension attaches Catch22 summarizations to each data-trend constituent, refines shared dataless clauses into dataful ones when a decision tree can separate class payloads, and scores each trace against the resulting clause set to build an embedding for a final white-box classifier. Across the four benchmarks, this produces accuracy, precision, recall, and F1 measures that are at least comparable to strong MTSC baselines and are better than the earlier dataless EMeriTAte on three datasets out of four; on the Dyskinetic events dataset the method reaches 100% accuracy where the best competitor stays at 25.71%. The paper also shows that loading and indexing the data once, then jointly mining all class-segmented logs, makes the dataful version faster than the previous two-phase approach.","pith_inferences":["The pipeline is not tied to the discretization source: the polyadic log and Poly-DECLARE machinery could be applied directly to genuinely event-based data (sensor logs, process logs with concurrent activities), since the paper's own conclusion says validating on event-based data is still open.","The decision-tree refinement learns only unary predicates on activation or target payloads; allowing dyadic predicates that compare activation and target payloads (e.g., 'B's value must exceed A's') would likely help classes whose discrimination is relational, at the cost of the oblique-tree overhead the paper mentions.","The performance gap on Italy Power Demand is a natural probe: re-enabling the pruned high-volatility patterns in the mining step and re-measuring would directly test whether the pruning rule is safe in general or whether it silently drops class-specific signal.","Because the final explanation is a propositional formula over clauses, the same machinery yields local explanations for misclassifications: a misclassified segment can be traced to which refined clauses it violated, which is a testable use for clinical monitoring."],"forward_implications":["If the reported accuracy holds across settings, explainable MTS classification no longer costs accuracy: class descriptions are propositional formulas over temporal clauses, and on the Dyskinetic dataset the same model that reaches 100% accuracy explains itself purely through cross-dimension correlations.","Because the polyadic log model allows multiple durative events at one timestamp, temporal correlations across dimensions are expressible as DECLARE-style constraints instead of being flattened into kernel values or attention weights, which is why the paper's method succeeds where competing classifiers score below 26% on the clinical data.","Refining shared dataless clauses into dataful ones only when a decision tree can separate payloads by class means accuracy can only improve over the dataless variant or stay equal; on the tested datasets the refinement is always on activation conditions, never on targets.","Jointly mining and embedding all class-segmented logs in one pass removes the separate trace-embedding phase of the previous EMeriTAte algorithm, which is what makes the dataful version faster despite computing additional payload statistics.","The method also works on univariate time series (Italy Power Demand, OSULeaf), so it is not limited to cross-dimensional problems, although on Italy Power Demand it lost accuracy because discarded high-volatility patterns were class-relevant."],"supporting_citations":[{"why":"Previous EMeriTAte solution and the Dyskinetic dataset; the baseline that the dataful variant extends and outperforms.","marker":"[14]"},{"why":"Supplies the data-trend (DT) discretization into durative constituents that the a priori phase extends with Catch22 payloads.","marker":"[20]"},{"why":"Defines the Catch22 time-series characteristics used as numerical payloads on mined constituents.","marker":"[25]"},{"why":"Bolt2 specification mining algorithm, the base that is extended to polyadic, concurrent, dataful mining.","marker":"[9]"},{"why":"Prior deviance model learning over declarative patterns, whose embedding and refinement approach this paper generalizes with white-box classifiers.","marker":"[12]"},{"why":"ROCKET convolutional-kernel classifier, a state-of-the-art MTSC baseline the paper compares against.","marker":"[16]"},{"why":"Canonical Interval Forest classifier using Catch22 in fixed windows, a competitor and conceptual contrast to DT-mined intervals.","marker":"[26]"},{"why":"TapNet attentional prototypical network, a strong baseline that fails on the clinical dataset where the proposed method reaches 100% accuracy.","marker":"[32]"}],"fun_headline_variants":["Explainable rules from sensor data beat state-of-the-art classifiers","Event-based rules make time series classification clear and accurate","Hybrid approach mines time series into readable rules with top accuracy","Closing the gap between numerical and event-based temporal classification","Poly-DECLARE rules explain every time series classification decision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the chosen discretization—pruning some high-volatility patterns and summarizing each remaining trend with Catch22 features—still contains the information that distinguishes the classes, so if the pruning removes class-relevant patterns the whole pipeline loses accuracy and the general superiority claim no longer holds.","fun_headline_variants_meta":{"raw":{"variants":["Explainable rules from sensor data beat state-of-the-art classifiers","Event-based rules make time series classification clear and accurate","Hybrid approach mines time series into readable rules with top accuracy","Closing the gap between numerical and event-based temporal classification","Poly-DECLARE rules explain every time series classification decision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000916,"raw_usage":{"total_tokens":3891,"prompt_tokens":863,"completion_tokens":3028,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2946}},"tokens_in":479,"tokens_out":3028,"duration_ms":21461,"temperature":1.0,"reasoning_tokens":2946,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:41:26.982711+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EMeriTAte+DF on Italy Power Demand with the currently discarded high-volatility patterns included in the DT-mining step: if accuracy returns to the roughly 99.6% level of the dataless EMeriTAte while runtime stays competitive, then the pruning rule, not the dataful refinement, is what limits this dataset, and the general claim that the pipeline retains all class-discriminative information is falsified.","supporting_citations":[{"cited_title":"In: Database Engineered Applications","cited_arxiv_id":null,"evidence_quote":"Previous EMeriTAte solution and the Dyskinetic dataset; the baseline that the dataful variant extends and outperforms."},{"cited_title":"Expert Systems with Applica- tions 201, 117176 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the data-trend (DT) discretization into durative constituents that the a priori phase extends with Catch22 payloads."},{"cited_title":"Data Min","cited_arxiv_id":null,"evidence_quote":"Defines the Catch22 time-series characteristics used as numerical payloads on mined constituents."},{"cited_title":"Computers 12(9) (2023)","cited_arxiv_id":null,"evidence_quote":"Bolt2 specification mining algorithm, the base that is extended to polyadic, concurrent, dataful mining."},{"cited_title":"Exploring Business Process Deviance with Sequential and Declarative Patterns","cited_arxiv_id":"2111.12454","evidence_quote":"Prior deviance model learning over declarative patterns, whose embedding and refinement approach this paper generalizes with white-box classifiers."},{"cited_title":"Data Min","cited_arxiv_id":null,"evidence_quote":"ROCKET convolutional-kernel classifier, a state-of-the-art MTSC baseline the paper compares against."},{"cited_title":"In: 2020 IEEE International Conference on Big Data (Big Data)","cited_arxiv_id":null,"evidence_quote":"Canonical Interval Forest classifier using Catch22 in fixed windows, a competitor and conceptual contrast to DT-mined intervals."},{"cited_title":"In: AAAI","cited_arxiv_id":null,"evidence_quote":"TapNet attentional prototypical network, a strong baseline that fails on the clinical dataset where the proposed method reaches 100% accuracy."}],"review_version":1}