{"id":"fc6d14c4-c7fe-4f6e-bbee-e1ef4c2e7990","arxiv_id":"2608.11704","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A DTW-based granular ball classifier, with random or label-informed ball splitting, improves robustness to symmetric label noise and cuts inference comparisons relative to DTW 1-NN on four benchmarks.","lead":"This paper groups similar time-series training examples into 'granular balls' and classifies new examples by comparing with ball centers instead of every stored example, which makes the classifier less vulnerable to mislabeled training data and reduces the number of distance computations at test time. The method is a simple modification of the standard nearest-neighbor time-series classifier, and the paper claims it is more robust to label noise and faster during inference.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by an unspecified rho-selection protocol: Section 5.2 states only that rho is 'searched', not that it is chosen by cross-validation, while Tables 3-4 and Figure 2 report 'best-performing' settings.","rationale":"The reader correctly identifies Eq. (7) as a heuristic that lacks formal justification in non-Euclidean DTW space, and that is worth testing. However, the most load-bearing concern for the paper's central claim is the model-selection protocol for rho. The claim that DTW-GBC 'generally mitigates' label-noise degradation is supported only by the experimental tables, and those tables may reflect selection over rho on the test set. If so, the results are optimistic upper bounds rather than evidence for a reproducible method. Even under a charitable reading, the paper does not state which rho values were selected for each dataset, how the ten noise draws interact with rho selection, or whether the reported variances include the model-selection step. The self-acknowledged limitation about kappa tuning on noisy labels reinforces this concern because the same issue applies to rho. A concrete rerun with an explicit nested cross-validation protocol would settle whether the claimed gains survive an unbiased selection rule. Until that check is performed, the conditional verdict is appropriate: the method is plausible and the experiments are suggestive, but the quantitative magnitude of the claimed advantage is not yet verified.","tokens_in":9513,"tokens_out":14185,"duration_ms":153774,"concrete_test":"Recompute Tables 3 and 4 with a pre-specified protocol: for each of the ten noisy training draws, select rho (and kappa for DTW-GBC-R) by five-fold cross-validation on the noisy training labels only, using the same candidate grids, then evaluate the selected model on the clean test set. Report the mean, standard deviation, and the selected rho values. If the cross-validated accuracy is materially lower than the table entries, the central claim's magnitude is not supported; if it matches, the ambiguity is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim rests entirely on Tables 3 and 4, yet the model-selection protocol is under-specified in a way that can directly bias these numbers. Section 5.2 says 'we fixed beta to 1 for two DTW-GBC based methods and searched rho in {0.5,0.6,0.7,0.8,0.9,1}'. The only sentence mentioning cross-validation is 'For DTW-GBC-R, the candidate values of kappa ... and the optimal value was selected using five-fold cross-validation on the training set.' It is therefore unclear whether rho is selected by cross-validation, by a fixed rule, or by test-set performance. Section 5.4 then reports 'best performances', and Figure 2 is captioned as 'corresponding to the best performances ... reported in Tables 3 and 4'. If Tables 3-4 pick, per dataset and noise level, the rho that maximizes test accuracy, the claimed accuracy gains are upper bounds over the rho grid rather than expected performance under a reproducible model-selection rule. The limitation paragraph in Section 6 concedes that tuning kappa on noisy training labels is fragile; the same fragility applies a fortiori to rho, because the cross-validation labels themselves are corrupted. Thus the robustness advantage over DTW 1-NN could shrink or vanish under a pre-registered protocol. This ambiguity is more load-bearing than the heuristic form of Eq. (7), because it directly affects the numerical support for the paper's headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DTW-based Granular Ball Computing (DTW-GBC), a method that constructs granular balls of training time series directly in the DTW distance space and classifies test samples by 1-NN over these balls. Two splitting strategies are introduced: random splitting (DTW-GBC-R) and label-informed splitting (DTW-GBC-L). Experiments on four benchmark datasets under symmetric label noise at rates 0.1 and 0.2 report that both variants consistently outperform vanilla DTW 1-NN in accuracy, weighted F1, and weighted G-mean, while requiring substantially fewer DTW comparisons at inference. The paper also provides a complexity analysis showing inference cost proportional to the number of granular balls rather than the number of training samples.","tokens_in":9842,"tokens_out":4081,"duration_ms":47293,"significance":"If the empirical claims hold, the paper offers a conceptually simple way to reduce the sensitivity of DTW-based nearest-neighbor classifiers to label noise and to lower inference cost, with a clean complexity analysis and evaluation on four datasets from two archives. The multiple metrics and ten noisy-label replications are appropriate, and the explicit acknowledgment of the fragility of hyperparameter selection on noisy labels is commendable. However, the significance is currently limited by the narrow baseline set, the unsupported distance-to-ball formula, and an underspecified hyperparameter-selection protocol that directly affects the reported numerical gains.","major_comments":[{"comment":"The hyperparameter selection protocol for rho is not specified. The text states only that rho was 'searched' over {0.5,0.6,0.7,0.8,0.9,1} and that Tables 3-4 report 'best performances', while Figure 2 is captioned as 'corresponding to the best performances'. If rho (or any other hyperparameter) was selected by test-set performance per dataset and noise level, the reported accuracy gains are upper bounds over the rho grid rather than expected performance under a reproducible model-selection rule. Please specify exactly how rho is chosen: for example, five-fold cross-validation on the training set as done for kappa, or a fixed rule. The limitation paragraph in Section 6 concedes that tuning kappa on noisy labels is fragile; rho deserves the same explicit treatment.","section":"Section 5.2 and Section 5.4"},{"comment":"The distance from a test series to a granular ball is defined as the DTW distance to the ball medoid minus the mean radius. No justification is given that subtracting the radius preserves the relevance ordering of balls in DTW space, which is generally non-Euclidean and can violate metric properties. This formula is load-bearing because it determines which ball is selected for prediction. Please provide a theoretical justification or an empirical evaluation comparing Eq. (7) with alternatives such as medoid-only distance or average distance to all ball members, and report whether the conclusions in Tables 3-4 are sensitive to this choice.","section":"Section 3.3, Eq. (7)"},{"comment":"The paper's headline claim of robustness to noisy labels is evaluated only against vanilla DTW 1-NN. The related-work section cites modern noisy-label time-series classifiers such as CTW and Scale-teaching, but no such method is included in the experiments. To support the broader claim implied by the title, please add at least one recent noisy-label TSC baseline, or explicitly restrict the claim to 'more robust than DTW-based 1-NN'. Without this, the quantitative evidence does not distinguish the proposed method from a simple averaging effect over granular balls.","section":"Section 2.1 and Section 5.4"},{"comment":"Both splitting strategies use random center selection, but the experiments report standard deviations only over the ten label-noise realizations, not over the randomness in ball construction. Since the random strategy is a core contribution, the variability of the results due to the random choices is unknown. Please report results over multiple random seeds or fix the random seeds and state this explicitly, so that the reported means and standard deviations are reproducible.","section":"Algorithm 1 and Section 5.2"}],"minor_comments":[{"comment":"The heading 'Dynamic Time Wraping based Granular Ball Computing' contains a typo; it should be 'Warping'.","section":"Section 3 heading"},{"comment":"The maximum splitting depth D is introduced in the complexity analysis without a formal definition. Please state how it is bounded in practice, e.g., in terms of the number of leaves or the stopping criterion in Eq. (6).","section":"Section 4"},{"comment":"The figure legend uses '#Balls' without clarifying that this is the average number of generated granular balls over the ten noisy-label realizations; please state this explicitly in the caption or axis labels.","section":"Figure 3"},{"comment":"For JapaneseVowels, the '#Timesteps' entry is '7-29'; please clarify that sequences have variable lengths and state how DTW handles this, since the notation T(i) in Section 3 already permits variable length.","section":"Table 1"},{"comment":"The paper reports only the number of DTW comparisons as an efficiency measure. Reporting wall-clock training and inference times would make the efficiency claim more concrete, given the substantial pairwise-DTW training cost.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: granular balls built directly in DTW distance space, with a label-informed medoid splitting rule, plus a granule-level 1-NN rule that cuts inference comparisons substantially. The experiments are internally consistent—both variants beat vanilla DTW 1-NN on all four datasets at both noise levels, and the efficiency numbers are concrete. That descriptive claim holds.\n\nWhat the paper does well: the idea is simple, clearly explained, and plausible. The recursive splitting criterion (purity threshold, min size) is standard GBC machinery, but applying it to DTW distances rather than encoded features is a legitimate step forward for time-series work. The label-informed strategy shows a modest but repeatable advantage over random splitting in most settings, which is a useful empirical finding. The complexity analysis is honest: training is dominated by the O(N^2 T^2) pairwise DTW computation, and inference scales with the number of balls rather than training samples.\n\nSoft spots, in proportion. The weak baseline is the biggest one. The paper compares only against vanilla DTW 1-NN, while citing robust TSC methods (CTW, Scale-Teaching) in the related work without benchmarking them. That leaves the magnitude of the robustness claim unanchored; noisy-label TSC has a literature with stronger baselines. The stress-test concern about rho is legitimate. Section 5.2 says rho is 'searched' over a grid, but only kappa for DTW-GBC-R is explicitly cross-validated. Section 5.4 then reports 'best performances' and Figure 2 is captioned as corresponding to the best-performing settings. If rho was selected by test-set accuracy per dataset and noise level, the reported gains are optimistic upper bounds. The paper's own limitation paragraph concedes that kappa selection on noisy training labels is fragile; rho is at least as fragile, since the cross-validation labels themselves are corrupted. Eq. (7)—distance to a ball as DTW-to-medoid minus radius—is a heuristic and the paper doesn't justify why that ordering is faithful in non-Euclidean DTW space. That bothered the reader, and it is a fair point, but I think it is secondary to the tuning ambiguity: even if Eq. (7) is imperfect, the method still works as a consistent heuristic, whereas the numerical support for the headline claim depends directly on how rho was chosen. No code or data are released, which makes the comparison harder to audit.\n\nWho this is for: people working on time-series classification under label noise, or on granular-ball methods for non-Euclidean data. It deserves a serious referee; the core idea is novel enough and the results are promising, but the tuning protocol and baselines need to be fixed before the claims are credible. A good referee report should push for pre-registered or at least explicitly described rho selection (cross-validation on noisy training labels), comparison with at least one robust TSC baseline, and code release.\n\nMy recommendation: send it to peer review, but condition acceptance on addressing the model-selection ambiguity and strengthening the baselines. The paper is a reasonable workshop-to-conference submission as is; with those fixes it could be a solid journal paper.","headline":"A genuinely new GBC-in-DTW-space method with consistent but thinly supported gains; the rho-selection ambiguity is the load-bearing problem, not the geometry heuristic.","tokens_in":10331,"tokens_out":998,"would_cite":false,"duration_ms":12812,"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":"Replacing individual training series with DTW-space granular balls preserves accuracy under label noise and cuts inference comparisons.","keywords":["time series classification","dynamic time warping","granular ball computing","label noise","nearest neighbor","multivariate time series","classification robustness","inference efficiency"],"falsifier":"On clean-label data, count how often the ball chosen by Eq. (7) differs from the ball that contains the test series' actual DTW nearest neighbor; if those disagreements are frequent and line up with prediction errors, the radius-subtraction distance is not a faithful ball distance and the method's benefit would have to be attributed to something else.","tokens_in":9305,"feed_emoji":"📈","tokens_out":10196,"duration_ms":97264,"temperature":0.7,"pith_summary":"This paper tries to establish that granular-ball computing can be carried out directly in the Dynamic Time Warping distance space of raw time series, and that the resulting classifier keeps its accuracy when training labels are randomly corrupted. The practical payoff is a replacement for DTW-based 1-nearest-neighbor classification, which is a strong baseline but inherits every mislabeled neighbor's class and must compare each test series against every training sample. DTW-GBC organizes the training set into a small number of granular balls and labels each test series by its nearest ball, so isolated bad labels are absorbed by the ball's majority vote and only one DTW comparison per ball is needed. On four benchmark datasets with 10% and 20% symmetric label noise, both proposed variants outperform plain DTW on accuracy, weighted F1, and weighted G-mean, while using roughly 38 to 190 comparisons per test sample instead of the full training-set size. If the result holds, it offers a cheaper default that tolerates bad labels for small-to-medium time-series classification collections.","feed_headline":"Two ball-based classifiers beat DTW 1-NN on noisy labels","feed_subtitle":"Tests at 10-20% label corruption keep accuracy while DTW comparisons drop up to 92 percent.","key_machinery":"The central object is the DTW-distance granular ball, a recursive cluster of training time series in the space defined by Dynamic Time Warping. Each ball carries three numbers: a medoid center (the series with smallest summed DTW distance to the others), a mean radius (average DTW distance of its members to that center), and a purity (fraction of members in the majority class). Balls are built by starting from the full training set and splitting any ball whose purity is below $\\rho$ and whose size is above $\\beta$ into $\\kappa$ children, with $\\kappa$ either random or derived from the class labels. The load-bearing mechanism is the ball-level scoring rule of Eq. (7), which turns each ball into one representative comparison point, and the purity threshold, which decides when further splitting is worth the extra inference cost.","core_discovery":"On its own terms, the paper's central claim is that the unit of nearest-neighbor classification for time series can be a granular ball rather than an individual series, and that this substitution simultaneously suppresses label noise and lowers inference cost. A ball is represented by a medoid center $C^{(g)}$, a mean radius $r^{(g)}$, and a purity equal to the majority-class fraction; a test series $X^{(i)}$ is scored against the ball by $\\mathrm{dist}(X^{(i)},\\mathrm{GB}^{(g)}) = d_{\\mathrm{DTW}}(X^{(i)},C^{(g)}) - r^{(g)}$, and the ball with the smallest score supplies the label. Balls are grown coarse-to-fine by recursively splitting any sufficiently large ball whose purity falls below a threshold, using either random child centers or a label-informed strategy that keeps the parent center and seeds one child per competing class. In the reported experiments, both variants beat conventional DTW on every dataset and metric at both noise levels, with the label-informed variant generally producing a more compact ball set. The same experiments show that test-time comparisons fall from the number of training series to the number of granular balls, a reduction of roughly one-third to over ninety percent.","pith_inferences":["Editorial inference: the paper's argument is not limited to its four datasets; the same construction should apply to any time-series benchmark where DTW is a competitive distance, and the natural next test is a sweep over the wider benchmark archives the four datasets were drawn from.","Editorial inference: the reported efficiency gain depends on the final number of balls $G$ staying well below the training size $N^{\\mathrm{tr}}$; on high-class-count or strongly impure datasets, a strict purity threshold could push $G$ close to $N^{\\mathrm{tr}}$ and erase the advantage, a worst case the paper does not quantify.","Editorial inference: replacing the random seeds in label-informed splitting with deterministic class medoids would make the method reproducible and probably more stable under high noise, but the paper only tests random seeds.","Editorial inference: if the radius-subtracted ball distance of Eq. (7) is a faithful ordering, the same distance could serve other time-series tasks such as clustering and anomaly scoring, which the paper lists as future work rather than evaluating."],"forward_implications":["Under symmetric label noise at rates 10% and 20%, DTW-GBC variants improve accuracy, weighted F1, and weighted G-mean over plain DTW on all four datasets tested.","Per test sample, inference requires one DTW comparison per granular ball rather than per training series; the experiments report reductions of about 33.5% to 92.4% in the number of comparisons.","The label-informed splitting strategy generally uses fewer balls than random splitting while matching or exceeding classification quality, so class structure can be used to compress the representation.","Raising the purity threshold makes both variants generate more balls, but the gain in G-mean saturates or slightly reverses, implying an optimal granularity for the robustness-efficiency trade-off.","Because pairwise DTW computation dominates the one-time training cost, any speedup in computing pairwise DTW would shrink training time without changing the ball-level inference scheme."],"supporting_citations":[{"why":"It supplies the DTW recurrence and distance measure that all ball construction and inference distances are built on.","marker":"[6]"},{"why":"It introduces granular-ball computing as the representation paradigm this paper adapts to DTW space.","marker":"[7]"},{"why":"It provides the granular-ball classifier baseline whose robustness properties motivate granule-level labeling.","marker":"[8]"},{"why":"It motivates distance-based time-series classification as the family DTW-GBC belongs to and competes with.","marker":"[5]"},{"why":"It establishes DTW 1-NN as the standard strong baseline for time-series classification.","marker":"[9]"},{"why":"It supplies the two univariate benchmark datasets used in the experiments.","marker":"[17]"},{"why":"It supplies the two multivariate benchmark datasets used in the experiments.","marker":"[18]"},{"why":"It supplies the symmetric label-noise model used to corrupt training labels.","marker":"[19]"}],"fun_headline_variants":["Granular balls make DTW classification noise-proof and faster","Ball-based DTW: robust time-series classification with fewer comparisons","Noisy time-series labels? Granular balls keep DTW robust and fast","Granular balls beat DTW 1-NN on noisy time-series"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that subtracting the ball's average radius from the DTW distance to its center produces a correct ordering of which ball a test series belongs to; if DTW distance space is not Euclidean enough for that subtraction to be meaningful, test series can be sent to the wrong balls and the robustness gains would not be explained by the mechanism claimed.","fun_headline_variants_meta":{"raw":{"variants":["Granular balls make DTW classification noise-proof and faster","Ball-based DTW: robust time-series classification with fewer comparisons","Noisy time-series labels? Granular balls keep DTW robust and fast","Granular balls beat DTW 1-NN on noisy time-series"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2900,"prompt_tokens":930,"completion_tokens":1970,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":1895}},"tokens_in":546,"tokens_out":1970,"duration_ms":13954,"temperature":1.0,"reasoning_tokens":1895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:30:04.880680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On clean-label data, count how often the ball chosen by Eq. (7) differs from the ball that contains the test series' actual DTW nearest neighbor; if those disagreements are frequent and line up with prediction errors, the radius-subtraction distance is not a faithful ball distance and the method's benefit would have to be attributed to something else.","supporting_citations":[{"cited_title":"Granular ball computing classifiers for efficient, scalable and robust learning,","cited_arxiv_id":null,"evidence_quote":"It provides the granular-ball classifier baseline whose robustness properties motivate granule-level labeling."},{"cited_title":"A review on distance based time series classification,","cited_arxiv_id":null,"evidence_quote":"It motivates distance-based time-series classification as the family DTW-GBC belongs to and competes with."},{"cited_title":"The great time series classification bake off: a review and experimental evaluation of recent algorithmic advances,","cited_arxiv_id":null,"evidence_quote":"It establishes DTW 1-NN as the standard strong baseline for time-series classification."},{"cited_title":"The ucr time series archive,","cited_arxiv_id":null,"evidence_quote":"It supplies the two univariate benchmark datasets used in the experiments."},{"cited_title":"Confident learning: Estimating uncertainty in dataset labels,","cited_arxiv_id":null,"evidence_quote":"It supplies the symmetric label-noise model used to corrupt training labels."}],"review_version":1}