{"id":"41af11aa-b1f7-42f3-892f-f94ad1ba66d3","arxiv_id":"2505.02173","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RDPC, a weighted blend of a trimmed L1 distance over the largest element-wise differences and the Pearson correlation dissimilarity, beats standard baselines on synthetic trend-and-season datasets and clusters electricity users.","lead":"A new dissimilarity measure for time series blends the largest point-by-point gaps with the Pearson correlation, improving clustering of simulated data with trends, seasons, and peaks. It also groups Thai electricity customers into seven interpretable usage classes, though the winning settings were chosen on the same test cases used for comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hyperparameters (alpha, p) are tuned on the same M/MC datasets used for the headline comparison, so the Table 3 outperformance may reflect selection on the evaluation set rather than intrinsic superiority.","rationale":"The reader's weakest-assumption analysis identified the same load-bearing concern: the synthetic M/MC datasets are used both for hyperparameter selection and for the head-to-head comparison, so the headline accuracy advantage may be inflated by selection on the evaluation set. My reading of Section 4.1 confirms this: the authors state that p = 0.1 and alpha = 0.2 were chosen because they 'provide the best accuracy for groups M and MC' on those exact datasets, and Table 3 then uses these values to claim superiority. The paper provides no repeated simulations, error bars, or independent validation, so the reported margins cannot be distinguished from tuning noise. This is the single most load-bearing weakness because the abstract's central claim rests entirely on Table 3; if the concern lands, the contribution reduces to a plausible dissimilarity with unvalidated parameter sensitivity. The proposed check—refitting on tuning splits and reporting held-out accuracy across many replications—would settle whether the advantage persists. I do not see a more fundamental mathematical flaw that would change the verdict; the metric proof gap in Proposition 3.8 is real but secondary and reparable. Therefore the reader's CONDITIONAL verdict remains appropriate, with no adjustment needed.","tokens_in":11824,"tokens_out":3650,"duration_ms":51810,"concrete_test":"Generate, say, 50 independent replications of each M and MC dataset configuration (same n, K, T, and generation rules, with new random seeds). For each replication, separate the clusters into a tuning set and a held-out set, select (alpha, p) on the tuning set only, and then report mean and standard deviation of held-out clustering accuracy for RDPC versus correlation-HC, DTW, GAK, and K-means. If RDPC's mean held-out accuracy is not above the best competitor by more than one standard deviation on a majority of replications, then the Table 3 outperformance is largely an artifact of tuning on the evaluation set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RDPC 'outperforms others in complicated cases involving different seasonal patterns, trends, and peaks.' The evidence for this claim is Table 3, where RDPC is compared with correlation-based hierarchical clustering, DTW, GAK, and K-means on groups M and MC. However, Section 4.1 explicitly states that the parameter combination p = 0.1 and alpha = 0.2 'provides the best accuracy for groups M and MC; thus, we use these parameter values throughout the remainder of the work.' This means the hyperparameters were selected on the very datasets that are later used to demonstrate outperformance, with no held-out set, no cross-validation, and no repeated simulations. With a single realization per dataset and a search over 30 parameter combinations, the reported accuracy values in Table 3 are optimistically biased. The generation process for M and MC is described only vaguely ('each cluster ... exhibits trends, seasonal patterns, and peaks at different times'), so it is unclear whether the selected parameters exploit idiosyncratic features of the specific simulated instances or a general property of the data family. If the tuning does not transfer to fresh datasets from the same generative process, the headline claim is an artifact of evaluation-set selection. A secondary, smaller correctness issue is that the proof of Proposition 3.8 contains a false inequality for the alpha = 1 case: it claims R_{i_j} <= R^xz_(n-j+1), which need not hold for an arbitrary index i_j; the intended triangle inequality for sums of largest order statistics can be proved differently, so this does not affect the main experimental claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new time-series dissimilarity, RDPC, defined as a convex combination of a ranked-differences term (a weighted average of a chosen fraction of the largest element-wise absolute differences) and the Pearson correlation dissimilarity. The authors integrate RDPC into hierarchical clustering and compare it with correlation-based hierarchical clustering, DTW, GAK, and K-means on nine simulated datasets, reporting superior accuracy on the four more complicated datasets (groups M and MC). They then apply the method to cluster 1,017 regular electricity users from a Thai utility into seven interpretable groups. The theoretical section proves nonnegativity, identity of indiscernibles, symmetry, and the triangle inequality under certain conditions.","tokens_in":12131,"tokens_out":7452,"duration_ms":93921,"significance":"If the empirical claims are validated, RDPC would be a useful and simple addition to the time-series clustering toolbox because it explicitly balances level information (absolute differences) with shape information (correlation), and the paper demonstrates interpretable clusters in a real electricity-consumption dataset. The construction is transparent, and the proofs for the uniform-weight case are mostly straightforward. However, no code or data are provided, and the main empirical evidence is weakened by the parameter-selection protocol, so the practical significance currently rests on a single favorable comparison.","major_comments":[{"comment":"The headline claim that RDPC outperforms other methods on groups M and MC is not supported by the experimental protocol, because the parameters p and alpha are selected on the same datasets later used for comparison. The text states that p=0.1 and alpha=0.2 \"provides the best accuracy for groups M and MC; thus, we use these parameter values throughout the remainder of the work,\" and Table 3 then reports RDPC accuracies of 0.9231, 1.0000, 1.0000, 0.9500, and 0.9286 on exactly those datasets. With one simulated instance per dataset and 30 parameter combinations searched, the reported superiority is optimistically biased and cannot be interpreted as an intrinsic advantage. Please provide independent validation, for example by drawing fresh datasets from the same generators, using a train/test split for parameter selection, and reporting accuracy variability over multiple simulations.","section":"Section 4.1 and Table 3"},{"comment":"Proposition 3.8 is stated for the general weights w in Definition 3.1, but its proof is valid only for uniform weights. The step sum_j w_j R^xz_{i_j} <= sum_j w_j R^xz_(n-j+1) does not follow for arbitrary positive weights because the indices i_j are ordered by the x-y differences, not by the x-z differences; for uniform weights the inequality is true because the sum over any r indices is bounded by the sum of the r largest values. Please restrict the proposition to the uniform-weight setting actually used in the paper, or provide a correct proof for general weights.","section":"Proposition 3.8"},{"comment":"The claim that M2 holds \"only for alpha>0\" is false when p=0. In that case r=0, RankDiff is identically zero, and d_RDPC reduces to (1-alpha)d_P for alpha<1, which violates M2 for non-identical perfectly correlated vectors; for alpha=1, d_RDPC is identically zero. The statement should explicitly require p>0, or Definition 3.1 should be adjusted. The applications use p=0.1, so this is a statement-level error rather than a flaw in the empirical results, but it must be corrected.","section":"Proposition 3.6"}],"minor_comments":[{"comment":"The weight vector w=(w1,...,wr) is said to belong to R^n, but it has only r components; it should be R^r. The empty sum arising when p=0 should also be explicitly defined.","section":"Definition 3.1"},{"comment":"The global alignment kernel reference appears as \"[?]\" in the text; it should be the numbered reference to Cuturi [22].","section":"Section 2.3.3"},{"comment":"Axiom (M4) is written as an equality d(x,y)=d(x,z)+d(z,y), but the standard triangle inequality is d(x,z) <= d(x,y)+d(y,z), and the proof of Proposition 3.8 establishes the inequality version. Please correct the axiom statement.","section":"Definition 3.4"},{"comment":"The term \"accuracy\" is not defined in the paper. Please specify how accuracy is computed (for example, as the proportion of correctly labeled series after matching cluster labels) and, ideally, report the variability across multiple simulation runs.","section":"Section 4.2 / Table 3"},{"comment":"The sentence \"This group is similar to RCM but more complicated\" appears to contain a typo: there is no RCM group defined, and it should presumably read \"similar to group M but more complicated.\"","section":"Section 4, group MC description"},{"comment":"The choice of seven clusters is justified by a subjective reading of the elbow plot (\"four distinct elbow points ... another slight elbow at nine\"). A cluster validity index or a more objective selection rule would make the real-data demonstration more convincing.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the circular parameter selection in Section 4.1: the method is tuned on the very datasets used for the headline comparison, so the reported outperformance is not yet a reliable result. The theoretical gaps (Propositions 3.6 and 3.8) are fixable by adding the p>0 assumption and restricting the triangle-inequality claim to uniform weights. If the authors can provide an independent evaluation with fresh simulations and report variability, the paper could become a useful contribution to applied time-series clustering. I would not accept the paper in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"At bottom this is a small but honest method paper: the RDPC dissimilarity, a convex combination of a trimmed L1 over the largest absolute differences and 1 - Pearson correlation, is new as a specific blend, and the basic metric properties are correctly proved. The idea is intuitive and easy to implement, and the electricity application is the kind of setting where both level and shape matter, so the motivation is genuine. If the claims hold up, practitioners in load profiling get a useful additional tool.\n\nWhere it falls down is the experimental evaluation. The hyperparameters alpha=0.2 and p=0.1 were chosen by scanning 30 combinations on the exact M and MC datasets that later appear in Table 3 as evidence of outperformance. That is selection on the evaluation set. With one simulated realization per dataset and no repeated runs or error bars, the reported accuracies are optimistic. The generative process for the M and MC groups is described only vaguely, so it is hard to tell whether the tuned parameters capture a general property of the family or idiosyncrasies of the particular instances. This is a load-bearing flaw for the abstract's claim of 'outperforms others in complicated cases.' The real-data section adds a further circularity: the same RDPC method is used to remove high-usage outliers before the final clustering, and the number of clusters K=7 is chosen by a subjective reading of elbow points.\n\nThe mathematics is mostly fine. Proposition 3.8's alpha=1 triangle inequality for RankDiff is actually proved correctly; the stress-test note's worry about a false inequality seems to misread the argument, which uses dominance by the largest order statistics rather than a termwise bound. Proposition 3.6 is correct for alpha>0, with the alpha=0 caveat appropriately called out. There are some editorial slips: missing references appear as '[?]' in Section 2.3.3 and the introduction, which a careful revision should fix.\n\nIn short: the measure itself is reasonable and the theory is sound, but the performance claim needs independent validation. I would send this to review because the method is plausible and the application is relevant, but the revision must address the tuning protocol — a separate validation set or cross-validation — and preferably repeated simulations. As it stands, the abstract overstates what Table 3 can support.","headline":"Simple new dissimilarity with sound metric properties, but the headline performance claim is weakened by tuning on the same datasets used for comparison.","tokens_in":12667,"tokens_out":3151,"would_cite":false,"duration_ms":37151,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"RDPC, a dissimilarity that mixes the largest pointwise differences with Pearson correlation, outperforms DTW, GAK, K-means, and correlation-based hierarchical clustering on time series with distinct seasonal patterns, trends, and peaks.","keywords":["time series clustering","dissimilarity measure","hierarchical clustering","Pearson correlation","ranked differences","electricity consumption","seasonal patterns","dynamic time warping"],"falsifier":"A direct test is to fix RDPC's parameters by cross-validation on held-out subsets of M and MC (or on a separate collection of labeled seasonal/trend/peak time series) and then measure clustering accuracy; if RDPC no longer outperforms correlation-based hierarchical clustering, the reported advantage is an artifact of tuning to the evaluation datasets. The paper itself supplies enough detail to re-run this comparison with the same generators and baselines.","tokens_in":11612,"feed_emoji":"⚡","tokens_out":15486,"duration_ms":135743,"temperature":0.7,"pith_summary":"The paper proposes RDPC, a dissimilarity measure for hierarchical clustering of time series, defined as a weighted combination of Pearson correlation dissimilarity and the average of a chosen fraction of the largest element-wise absolute differences. The measure is built for situations in which both the overall level of a series and its shape or association matter, as with electricity consumption. The paper's central claim is that on synthetic datasets containing different seasonal patterns, trends, and peaks, hierarchical clustering with RDPC recovers the true groups more accurately than correlation-based hierarchical clustering, DTW, GAK, and K-means. It then applies RDPC to a random sample of Thai electricity users and reports seven interpretable clusters, such as declining, increasing, winter-peaking, and 2022/2023 drop users.","feed_headline":"Blending distance and correlation wins on seasonal time series","feed_subtitle":"RDPC adds pointwise gaps to Pearson correlation, beating DTW and K-means on hard synthetic and real power-use data.","key_machinery":"The machinery is the ranked-differences Pearson correlation dissimilarity, $d_{\\mathrm{RDPC}}(x,y|\\alpha,p,\\mathbf{w}) = \\alpha\\,\\mathrm{RankDiff}(x,y|p,\\mathbf{w}) + (1-\\alpha)\\,d_P(x,y)$, where $d_P(x,y)=1-\\mathrm{Corr}(x,y)$ and $\\mathrm{RankDiff}$ is the sum $\\sum_{j=1}^{r} w_j R_{(n-j+1)}$ of the $r=\\lceil pn\\rceil$ largest absolute element-wise differences, ordered and weighted by a probability vector $\\mathbf{w}$. With uniform weights, $\\mathrm{RankDiff}$ is just the mean of the largest $r$ gaps, so the parameter $\\alpha$ controls how much the clustering leans on level differences versus shape correlation, and $p$ controls how many of the largest gaps are considered. This object does the argument's work: it lets a single hierarchical-clustering run separate series that share a level but differ in pattern, or share a pattern but differ in level.","core_discovery":"The central discovery is that a simple convex combination of two classical notions—shape similarity via $1-\\mathrm{Corr}(x,y)$ and level similarity via the average of the largest $\\lceil pn\\rceil$ absolute differences—produces a dissimilarity that is simultaneously sensitive to both, and that this combined measure, when used with complete-linkage hierarchical clustering, is more accurate than either perspective alone on the hard synthetic datasets. With the tuning chosen in the paper ($\\alpha=0.2$, $p=0.1$, uniform weights), RDPC attains accuracies 0.9231, 1.0000, 1.0000, 0.9500, and 0.9286 on the M1, M2, M3, MC1, and MC2 datasets, higher than the corresponding accuracies of correlation-based hierarchical clustering, DTW, GAK, and K-means. The paper also shows that RDPC is a symmetric, non-negative function that is a metric in the limiting case $\\alpha=1$ and that, for $\\alpha=0$, it converges almost surely to unity for independent series.","pith_inferences":["The same ranked-differences construction could be paired with Spearman or Kendall correlation instead of Pearson, which would make the shape component robust to outliers and monotone rescaling while keeping the level-sensitive term.","The parameter $\\alpha$ could plausibly be set per dataset by an internal cluster-validity index rather than fixed at 0.2; the paper's own sensitivity table shows accuracy varies with $\\alpha$, so a data-driven choice might generalize better to unseen data.","The iterative two-group splitting used to remove the 157 high-usage users can be read as a standalone outlier-screening procedure; it could be tested as a general preprocessing step for skewed consumption data."],"forward_implications":["A single dissimilarity can separate time series that differ in level, shape, or both, so hierarchical clustering with RDPC applies directly to domains like energy, water, and healthcare where raw scale carries meaning.","On the five complicated synthetic datasets, the reported accuracy advantage over DTW and GAK suggests that time-warping alignment is the wrong tool when seasonality must be preserved, since warping January to April destroys the very pattern that defines a group.","The elbow method, applied with RDPC, recovers the true number of clusters for the M and MC datasets, whereas the baselines' elbow points do not align with the true labels, indicating the measure gives usable cluster-count diagnostics in complicated settings.","The application to Thai electricity-consumption data identifies small but interpretable consumer groups, such as winter users and users whose usage dropped sharply in a specific year, that would be missed by clustering on level or on correlation alone."],"supporting_citations":[{"why":"Shows that Euclidean distance on scaled data equals Pearson correlation dissimilarity, motivating a measure that keeps raw scale when scaling is inappropriate.","marker":"[20]"},{"why":"Supplies the dynamic time warping baseline that RDPC must beat on seasonal data.","marker":"[9]"},{"why":"Supplies the global alignment kernel baseline used in the accuracy comparison.","marker":"[22]"},{"why":"Supplies the K-means baseline that clusters by Euclidean distance in the comparison.","marker":"[4]"},{"why":"Provides hierarchical clustering, the framework into which RDPC is plugged with complete linkage.","marker":"[5]"},{"why":"Provides the elbow method used to detect the number of clusters in the experiments and the application.","marker":"[21]"}],"fun_headline_variants":["Hybrid distance-correlation dissimilarity tops seasonal benchmarks","RDPC merges gap and correlation to outcluster DTW and K-means","Ranked differences plus Pearson correlation: better time-series clusters","New dissimilarity blends peak gaps and shape for winning clusters","Combining level and shape similarity improves clustering accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic datasets M and MC, together with the parameter values ($\\alpha=0.2$, $p=0.1$) chosen because they maximize accuracy on exactly those datasets, faithfully represent the real-world “complicated cases” the method claims to handle.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid distance-correlation dissimilarity tops seasonal benchmarks","RDPC merges gap and correlation to outcluster DTW and K-means","Ranked differences plus Pearson correlation: better time-series clusters","New dissimilarity blends peak gaps and shape for winning clusters","Combining level and shape similarity improves clustering accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000964,"raw_usage":{"total_tokens":4103,"prompt_tokens":943,"completion_tokens":3160,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":3076}},"tokens_in":559,"tokens_out":3160,"duration_ms":26831,"temperature":1.0,"reasoning_tokens":3076,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:59:40.038425+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to fix RDPC's parameters by cross-validation on held-out subsets of M and MC (or on a separate collection of labeled seasonal/trend/peak time series) and then measure clustering accuracy; if RDPC no longer outperforms correlation-based hierarchical clustering, the reported advantage is an artifact of tuning to the evaluation datasets. The paper itself supplies enough detail to re-run this comparison with the same generators and baselines.","supporting_citations":[{"cited_title":"Fast global alignment kernels,","cited_arxiv_id":null,"evidence_quote":"Supplies the global alignment kernel baseline used in the accuracy comparison."},{"cited_title":"Some methods for classification and analysis of multivariate observations,","cited_arxiv_id":null,"evidence_quote":"Supplies the K-means baseline that clusters by Euclidean distance in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides hierarchical clustering, the framework into which RDPC is plugged with complete linkage."},{"cited_title":"Boehmke and B","cited_arxiv_id":null,"evidence_quote":"Provides the elbow method used to detect the number of clusters in the experiments and the application."}],"review_version":1}