{"id":"b3176652-5c2c-4c6a-a44d-de44b00e1bb5","arxiv_id":"2504.14206","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"TransDe combines HP-filter decomposition, multi-scale patch transformers, and stop-gradient KL contrastive learning to score multivariate time series anomalies, reporting top F1 on four of five public benchmarks.","lead":"This paper proposes TransDe, a transformer-based anomaly detector that splits time series into trend and cycle components, patches them at multiple scales, and uses KL divergence between two attention views as the anomaly score. It reports F1 gains over twelve baselines on four of five public datasets, but underperforms on SMD and omits several evaluation details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"As written, the training loss in Eq. 12 is identically zero because Eqs. 10–11 define the same symmetric KL sum, so the described method cannot learn or produce the reported F1 scores.","rationale":"The reader's weakest_assumption targets the absence of a statistical justification for KL divergence as an anomaly score. That is a real issue, but it is not the most load-bearing point: an unproven score could still be validated empirically, whereas an identically zero training loss cannot train the model at all. The internal inconsistency in Section 4.5 was noted in the reader's rationale but was not chosen as the weakest assumption. I agree partially: both concerns undermine the preprint, but the zero-loss collapse is decisive as written. The attention dimension mismatches in Eqs. (2)–(6) compound the incoherence, and the contradiction between the 'two encoder layers' in Section 5.4 and the 'optimal number is 3' in Section 5.8 further disconnects the written configuration from the reported experiments. If the repository contains a different loss, the paper needs a major revision to describe the actual method; if the repository contains the same equations, the experimental results cannot be reproduced from the stated method. Either way, the abstract's claim of superiority is not supported by the written manuscript, so the reader's REJECT verdict should stand unchanged.","tokens_in":17444,"tokens_out":4182,"duration_ms":36139,"concrete_test":"Check the released repository by printing L_intra and L_inter during a training step and verify whether L_intra − L_inter is, as written, zero for all inputs. Also re-derive Eq. (12) from Eqs. (10)–(11) on a simple two-element example; if the paper's formulas are followed literally, the difference is identically zero. If the code instead contains an asymmetric or otherwise different loss, compare it line-by-line with Eqs. (10)–(12) and document the discrepancy. If the code matches the paper, the experimental results cannot be reproduced from the stated method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is not only the unproven KL premise; it is that the training objective written in Section 4.5 is identically zero. Equation (10) is L_intra = Σ [KL(Ω_intra ‖ Stop(Ω_inter)) + KL(Stop(Ω_inter) ‖ Ω_intra)]. Equation (11) is L_inter = Σ [KL(Ω_inter ‖ Stop(Ω_intra)) + KL(Stop(Ω_intra) ‖ Ω_inter)]. Since addition is commutative, both expressions are exactly the same sum, so L_intra = L_inter for every possible input. Equation (12), L = (L_intra − L_inter)/C, is therefore zero for every input. A loss that is identically zero cannot train the encoder; the pipeline described in the paper has no learning signal, and the reported F1 values cannot be produced by the method as specified. The dimension mismatches in Equations (2)–(6), e.g., Q_intra = W_Q X_intra with W_Q ∈ R^{(d_model/h)×(d_model/h)} and X_intra ∈ R^{P×d_model} cannot yield Q_intra ∈ R^{N×d_model/h}, reinforce that the written equations do not describe a coherent model. The central claim survives only if the released repository implements a different objective than Eq. (12); in that case the paper does not actually describe the method that generated the numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TransDe, an unsupervised multivariate time series anomaly detection framework that combines Hodrick-Prescott decomposition, multi-scale patch-based transformer encoders, and a KL-divergence contrastive objective with stop-gradient. The authors claim that decomposing the series into trend and cyclical components and learning inter-patch and intra-patch dependencies through contrastive learning yields state-of-the-art F1 scores on five public benchmarks (SMD, MSL, SMAP, SWaT, PSM) compared with twelve baselines. The method is described in Section 4, with experiments in Section 5 that report best F1 values of 87.33, 94.36, 96.67, 98.04, and 96.43 on the five datasets respectively.","tokens_in":17651,"tokens_out":9443,"duration_ms":75306,"significance":"If the proposed framework worked as described, it would be a competitive contribution to time series anomaly detection, particularly in its use of decomposition and contrastive learning without negative samples. The paper includes ablations for stop-gradient, patch-level choices, and loss variants, and it provides a public code link. However, the central methodological derivation contains a degenerate training objective and multiple dimension inconsistencies that make the described model incoherent. The empirical results, while strong on some datasets, are not supported by the written method, and the absence of threshold-selection details and error bars further weakens the evaluation. The core idea is plausible, but the current manuscript does not provide a reproducible or internally consistent method.","major_comments":[{"comment":"As written, L_intra and L_inter are the same sum of two KL terms up to commutation: L_intra = KL(Ω̂_intra||Stop(Ω̂_inter)) + KL(Stop(Ω̂_inter)||Ω̂_intra) and L_inter = KL(Ω̂_inter||Stop(Ω̂_intra)) + KL(Stop(Ω̂_intra)||Ω̂_inter) evaluate identically in the forward pass because Stop(·) is the identity. Consequently L = (L_intra − L_inter)/C is identically zero for every input, so Eq. (12) does not define a learning signal. The stop-gradient only changes the backward pass; the paper neither acknowledges this nor explains why optimizing a zero-valued objective with a pseudo-gradient is valid. This is a load-bearing flaw because the training procedure described cannot produce the reported F1 values.","section":"4.5, Eq. (12)"},{"comment":"Several dimension mismatches make the attention equations incoherent. In Eq. (2), W_Q, W_K ∈ R^{(d_model/h)×(d_model/h)} are multiplied by X_intra ∈ R^{P×d_model}, but the product is claimed to be Q_intra, K_intra ∈ R^{N×d_model/h}; the inner dimensions d_model/h and d_model do not match, and the output length P is replaced by N. Similarly, Eq. (5) with X_inter ∈ R^{N×d_model} and the same W_Q cannot yield Q_inter ∈ R^{N×d_model/h}. Moreover, Eq. (3) would give Z_intra ∈ R^{N×N} (or P×P depending on the correct shapes), but Eq. (4) asserts Ω_intra ∈ R^{P×P}, with no explanation of how the head dimension and length are reconciled. These inconsistencies mean the encoder is not specified as a computational graph.","section":"4.3, Eqs. (2)–(7)"},{"comment":"The abstract claims superiority over twelve baselines in terms of F1 score, but Table 4 shows that on SMD, TransDe (F1 = 87.33) is outperformed by NormFAAE (93.82), TranAD (91.57), DCFF-MTAD (91.40), and AnomalyTrans (90.33). The text acknowledges this only later in Section 5.5, which conflicts with the unqualified claim in the abstract and in the contributions list.","section":"5.5, Table 4 and Abstract"},{"comment":"The anomaly threshold ρ in Eq. (14) is a free hyperparameter, yet the paper does not describe any procedure for setting it (e.g., on a validation set, by quantile, or by grid search). If ρ is selected using test-set labels, the reported F1 values are optimistically biased and cannot be compared with baselines that may have used different threshold strategies. The manuscript should state the threshold selection rule and, ideally, report the dependence of F1 on ρ.","section":"4.6 and 5.4"},{"comment":"The central assumption that anomalies cause a high KL divergence between Ω̂_intra and Ω̂_inter is asserted without derivation or citation: 'any anomalies with rare patterns will result in inconsistent representations, leading to a high anomaly score.' No statistical or empirical argument is given for why normal data must yield small divergence, and because the training loss in Eq. (12) is degenerate, the paper provides no learned-encoder justification for this premise.","section":"4.6"}],"minor_comments":[{"comment":"Section 5.4 says TransDe includes two encoder layers, but Section 5.8 concludes the optimal number of layers is 3; this inconsistency affects reproducibility and should be resolved.","section":"5.4 and 5.8"},{"comment":"Table 4 contains typographical errors: '0.86.37' for OmniAnomaly on MSL and '0.92.45' for IForest on PSM; the SMD column for IForest shows '76.09 0.92.45 83.48' which should likely be '76.09 92.45 83.48'.","section":"Table 4"},{"comment":"The references contain duplicates: Su et al. 2019a and 2019b are the same paper, and Xu et al. 2022a and 2022b are the same paper.","section":"References"},{"comment":"No standard deviations or multiple seeds are reported, which makes the claimed improvements of 0.11% and 0.26% on MSL and SMAP hard to assess.","section":"Section 5.5"},{"comment":"The visualization in Section 5.9 uses a UCR univariate dataset without identifying which dataset is used or explaining why a multivariate model is evaluated on a univariate sequence.","section":"Section 5.9"}],"recommendation":"reject","confidential_remarks":"The paper's central methodological description is irreparably flawed as written: the training loss in Eq. (12) is degenerate and the attention equations in Section 4.3 are dimensionally inconsistent. These are not presentation issues but load-bearing errors that prevent the reader from reproducing or trusting the reported results. I concur with the reader's verdict. If the authors can present a coherent objective and correct the equations, a resubmission might be considered, but the current version cannot be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things up front. First, the paper builds a genuinely new combination — HP-filter decomposition, multi-scale patch transformers, and a stop-gradient symmetric-KL contrastive loss — and on four of five benchmarks it reports F1 gains over several strong baselines. Second, the write-up is not trustworthy as it stands: the equations don't type-check, the implementation contradicts the hyperparameter analysis, and the evaluation reporting is incomplete.\n\nThe strongest part is the experiment suite. Five standard datasets, twelve baselines, and ablations on stop-gradient, patch level, loss function, normalization, contrastive paradigm, and hyperparameters. The SMD weakness is acknowledged honestly, which is more than many papers do. If the code matches the intended design, this is a competitive incremental method.\n\nNow the soft spots, in order of severity. Equations (2)–(6) have dimension mismatches: a weight matrix in R^{(d_model/h)×(d_model/h)} multiplied by an embedding in R^{P×d_model} cannot produce Q in R^{N×d_model/h}. These need to be rewritten before anyone can verify the architecture. The stress-test claim that Eq. (12) is identically zero does not hold: L_intra and L_inter are symmetric KL divergences between different argument pairs, so their difference is not zero in general. But the loss is still odd — it subtracts two positive terms and is never clearly motivated — and the anomaly score in Eq. (13) is not the same as the training loss, which is unexplained. There are also concrete internal contradictions: Section 5.4 says two encoder layers while Section 5.8 says the optimum is three; Table 3's patch sizes for SMAP and SWaT differ from the \"best\" values in Table 8. No error bars are given, threshold selection is not described, and the per-dataset patch/window search looks like tuning on test data. The omission of DCdetector and TFAD is glaring since both are direct predecessors of pieces of this method.\n\nWho is this for? A reader working on contrastive or decomposition-based time series anomaly detection who wants to see one more engineering point on the F1 frontier. It is not a conceptual breakthrough. It deserves a serious referee because the empirical claim is sharp and the architecture is coherent enough to be salvageable, but the current text cannot be accepted as the description of the method that produced the numbers.\n\nRecommendation: send it to review, and tell the authors to fix the equations, reconcile the hyperparameters, and supply error bars and a threshold-selection protocol. If the code checks out, major revision could make it publishable.","headline":"Plausible incremental architecture with competitive F1 results on four of five benchmarks, but the written method is internally inconsistent and needs substantial revision before it can be evaluated.","tokens_in":18255,"tokens_out":4433,"would_cite":false,"duration_ms":38201,"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":"TransDe claims that decomposing a time series into trend and cycle, then aligning intra-patch and inter-patch transformer views with KL divergence, yields leading F1 scores on four multivariate anomaly benchmarks.","keywords":["time series anomaly detection","unsupervised learning","transformer","time series decomposition","contrastive learning","KL divergence","patch-based representation","multivariate time series"],"falsifier":"Train TransDe on normal data only, then feed a held-out normal segment containing legitimate abrupt changes such as sensor recalibration, mode switches, or level shifts, and count how often the KL score crosses the threshold $\\rho$; if many normal points score above the threshold, the underlying consistency assumption is false.","tokens_in":17150,"feed_emoji":"📊","tokens_out":9985,"duration_ms":78418,"temperature":0.7,"pith_summary":"TransDe, the framework proposed here, is an unsupervised multivariate time-series anomaly detector that combines three moves: Hodrick-Prescott decomposition into trend and cyclical components, multi-scale patching with a shared-weight transformer that computes intra-patch and inter-patch attention views, and a contrastive loss that aligns these views with symmetric KL divergence under a stop-gradient. The paper claims this combination learns normal patterns well enough that anomalies, which produce inconsistent views, receive high anomaly scores. On four public benchmarks (MSL, SMAP, SWaT, and PSM), the reported F1 scores are 94.36, 96.67, 98.04, and 96.43, ahead of twelve baselines; on SMD the method trails the best baseline, which the authors attribute to single-point anomalies and to normal points being misread as anomalous. The central assertion is that decomposition plus contrastive patch views, rather than reconstruction or prediction error, is the right objective for anomaly scoring under noise.","feed_headline":"Decomposed patch transformer tops 12 baselines in anomaly F1","feed_subtitle":"Splitting series into trend and cycle, then comparing patch views, lifts F1 on four public benchmarks.","key_machinery":"The load-bearing object is the pair of patch-level attention views, $\\hat{\\Omega}_{\\mathrm{intra}}$ and $\\hat{\\Omega}_{\\mathrm{inter}}$, produced from HP-filtered trend and cyclical components and expanded to the same $\\mathbb{R}^{T\\times T}$ dimension. The Hodrick-Prescott filter is a smoothing procedure that splits a series $x_t$ into a trend $\\tau_t$ and a cyclical residual $\\varepsilon_t$; this is the decomposition that feeds the multi-scale patch channels. The paper uses the symmetric KL divergence between the two views as a stand-in for pattern consistency: normal patterns are supposed to agree across views, so divergence stays low, while rare anomalous patterns are supposed to disagree, so divergence rises. A stop-gradient operation on the target view makes the loss asymmetric in update flow, which the paper claims prevents model collapse and improves the learned embeddings, and multi-scale patch sizes act as independent channels whose losses are combined.","core_discovery":"On its own terms, the paper's discovery is a detection pipeline in which the anomaly score is the KL divergence between two transformer-generated views of the same decomposed, patched sequence. The trend and cyclical components are segmented at several patch sizes; one attention pass looks inside patches (intra-patch dependencies, a $\\mathbb{R}^{P\\times P}$ matrix) and another looks across patches (inter-patch dependencies, a $\\mathbb{R}^{N\\times N}$ matrix). These representations are expanded to a common $\\mathbb{R}^{T\\times T}$ size, fused across components, and scored by the symmetric KL divergence between the two views, with a stop-gradient applied to the target side; the same loss is used for training and for the test-time anomaly score. The reported experiments give best F1 values of 94.36 on MSL, 96.67 on SMAP, 98.04 on SWaT, and 96.43 on PSM, with SMD at 87.33, and the authors attribute the gains to isolating trend and cyclical structure and to avoiding MSE-style optimization that is vulnerable to noise.","pith_inferences":["Editorial inference: the paper treats HP decomposition as fixed preprocessing, but the contrastive objective does not depend on the specific filter; swapping in other decomposition schemes would test whether decomposition itself or the HP choice carries the gain.","Editorial inference: the anomaly score is exactly the training loss with stop-gradient, so threshold selection may determine much of the reported F1, and the paper proposes no principled thresholding method.","Editorial inference: since one attention head and shallow encoders perform best, the transformer may be acting as a lightweight alignment module rather than a deep sequence model; replacing it with a linear or convolutional patch projector would be a natural testable extension.","Editorial inference: the SMD weakness and the false positives on normal segments shown in the visualization suggest the core assumption that rare patterns always inflate KL divergence may fail on point anomalies; per-channel thresholds or a prior on expected divergence are natural extensions."],"forward_implications":["The reported F1 results imply that reconstruction and prediction objectives are not required for strong anomaly detection; a divergence-based score computed directly from two attention views is enough to outperform twelve baselines on four datasets.","Because the loss is computed from positive pairs only, the framework avoids negative-sample mining, and the paper's timing comparison indicates lower training cost per epoch than SimCLR- and MoCo-style contrastive learning.","The ablations show that both inter-patch and intra-patch channels contribute: dropping either reduces F1, so the multi-scale two-view design, not the transformer alone, is doing the work.","The stop-gradient and symmetric-KL choices are load-bearing: ablations in which either is removed degrade performance across datasets.","The framework is not uniformly better: on SMD it falls below the best baseline, and the authors link this to single-point anomalies and to normal signals being misclassified."],"supporting_citations":[{"why":"Supplies the MSL and SMAP spacecraft benchmarks where TransDe reports its top F1 values.","marker":"Hundman et al., 2018"},{"why":"Supplies the SWaT water-treatment benchmark on which TransDe achieves 98.04 F1.","marker":"Mathur and Tippenhauer, 2016"},{"why":"Supplies the PSM server-metric benchmark used for the 96.43 F1 result.","marker":"Abdulaal et al., 2021"},{"why":"Supplies the SMD server benchmark, the dataset where TransDe underperforms the best baseline.","marker":"Su et al., 2019a"},{"why":"Provides the positive-pair-only contrastive loss with stop-gradient that TransDe adapts into its asynchronous KL objective.","marker":"Chen and He, 2021"},{"why":"The source cited for the Hodrick-Prescott decomposition that splits each input into trend and cyclical components.","marker":"Zhang et al., 2022"},{"why":"The strongest transformer baseline (AnomalyTransformer) whose F1 results define the comparison TransDe must beat.","marker":"Xu et al., 2022b"},{"why":"SimCLR, used in ablation to benchmark the no-negative-sample contrastive paradigm against negative-sample mining.","marker":"Chen et al., 2020"},{"why":"MoCo, the second negative-sample contrastive baseline in the ablation and training-time comparison.","marker":"He et al., 2020"}],"fun_headline_variants":["KL divergence on decomposed patches tops 12 anomaly baselines","Contrastive decomposed patches beat MSE for anomaly detection","Multi-scale transformer with KL score sets anomaly F1 records","Decomposition plus patch contrast lifts anomaly detection F1","Trend-cycle split and contrastive patches outperform MSE for anomaly F1"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire detection pipeline rests on the unproven assumption that every anomaly inflates the KL divergence between the two patch views while every normal segment keeps it low; if a normal but unusual regime or a subtle point anomaly violates this, the framework will either raise false alarms or miss anomalies.","fun_headline_variants_meta":{"raw":{"variants":["KL divergence on decomposed patches tops 12 anomaly baselines","Contrastive decomposed patches beat MSE for anomaly detection","Multi-scale transformer with KL score sets anomaly F1 records","Decomposition plus patch contrast lifts anomaly detection F1","Trend-cycle split and contrastive patches outperform MSE for anomaly F1"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00078,"raw_usage":{"total_tokens":3471,"prompt_tokens":992,"completion_tokens":2479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":2398}},"tokens_in":608,"tokens_out":2479,"duration_ms":15810,"temperature":1.0,"reasoning_tokens":2398,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:54:05.053504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train TransDe on normal data only, then feed a held-out normal segment containing legitimate abrupt changes such as sensor recalibration, mode switches, or level shifts, and count how often the KL score crosses the threshold $\\rho$; if many normal points score above the threshold, the underlying consistency assumption is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SWaT water-treatment benchmark on which TransDe achieves 98.04 F1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PSM server-metric benchmark used for the 96.43 F1 result."}],"review_version":1}