{"id":"36f69edd-6fdd-424e-9fe3-b96709d10334","arxiv_id":"2501.16364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MtsCID combines frequency-domain transformers, multi-scale patch attention, and fixed sinusoidal prototypes to detect anomalies, outperforming nine baselines on six of seven datasets.","lead":"A new detection model, MtsCID, learns coarse-grained normal patterns in both the time and cross-variable directions and reports top F1 scores on six of seven standard multivariate time series anomaly benchmarks. The gains over existing methods are modest on mature datasets but large on one harder benchmark, and the evaluation omits error bars and a direct comparison to one closely related frequency-domain method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The anomaly score (Eq. 17) relies on a relationship-deviation term that is never minimized during training; without evidence that RD separates normal from anomalous timestamps, the claimed inter-variate gains from sinusoidal prototypes are unsupported.","rationale":"The paper's claimed improvement over baselines rests on the dual-network design and, specifically, on the sinusoidal prototypes interaction module. The proposed check targets a more basic and more concrete issue than whether real-world inter-variate relationships lie near the sinusoidal subspace: even if those relationships were representable, the lower branch has no training objective that makes RD small or discriminative on normal data. Since the upper and lower branches are trained with separate losses and share no parameters, replacing RD with a constant at inference isolates the contribution of the relationship-deviation term without retraining. This is a clean empirical test of whether the prototype interaction is load-bearing. The reader's verdict is already conditional, and this concern adds a precise condition that should be verified before the central claim is accepted; therefore no change to the verdict is needed.","tokens_in":606,"tokens_out":6464,"duration_ms":176840,"concrete_test":"Using the released code, on each of the seven datasets compute anomaly scores with the same trained model under Eq. 17 and under Eq. 17 with RD replaced by a constant (so Softmax is uniform and the ranking reduces to TD alone). If point-adjustment F1 and VUS-ROC/VUS-PR change by less than 1%, the sinusoidal-prototype signal is not load-bearing; if they change materially, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central performance claim depends on Eq. 17, which multiplies temporal deviation TD by Softmax over relationship deviations RD(O_t,:, M_::). However, the only training signal for the lower branch is the entropy loss L_i_ent (Eq. 15). That loss encourages the softmax attention over the fixed sinusoidal prototypes to be peaky, but it never penalizes the distance between O and the nearest prototype. RD is therefore not optimized to be small, calibrated, or even meaningful on normal data; it can be arbitrarily large or unscaled, and the softmax then acts only on relative, uncalibrated values. Unless RD by itself separates normal from anomalous timestamps, Eq. 17 is simply a reweighted TD with no learned inter-variate contribution. The ablation in Table 4 removes the entire lower branch, which conflates the RD scoring mechanism with extra network capacity and does not isolate the effect of the prototype memory. The paper reports no RD distributions and no inference with RD removed, so the claimed boost from sinusoidal prototypes is not demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MtsCID, a semi-supervised multivariate time series anomaly detection method with a dual-network architecture. The upper branch (t-AutoEncoder) captures coarse-grained intra-variate temporal dependencies from attention maps of multi-scale patches in the frequency and time domains; the lower branch (i-Encoder) captures inter-variate relationships through convolution, a frequency-domain Transformer, and interaction with a fixed matrix of sinusoidal prototypes. Training combines a reconstruction loss with an entropy loss that encourages peaked prototype attention. At inference, the anomaly score combines temporal deviation with a softmax-weighted relationship deviation. The method is evaluated on seven public datasets against nine baselines, with additional ablation, sensitivity, and scalability analyses.","tokens_in":14418,"tokens_out":4480,"duration_ms":45637,"significance":"If the empirical claims are fully supported, MtsCID would be a useful addition to the MTS anomaly detection literature: the coarse-grained patch-attention idea is well motivated, the fixed sinusoidal prototype memory avoids the training instability of learned memory items, and the paper evaluates on multiple metrics and provides code. The paper also reports sensitivity analyses and a scalability study, which are valuable. However, the central comparative claim currently rests on a partially non-reproduced baseline comparison, threshold selection on test data, and an anomaly-score component whose discriminative value is not directly demonstrated.","major_comments":[{"comment":"The relationship deviation term RD in Eq. (17) is never explicitly minimized during training: the only lower-branch supervision is the entropy loss L_i_ent in Eq. (15), which encourages peaked weights but does not constrain the distance between O and the fixed sinusoidal prototypes. The paper reports no RD distributions, no experiment with RD removed from the score, and no comparison against random prototypes; the ablation in Table 4 removes the entire lower branch, which conflates the scoring mechanism with added network capacity. This is load-bearing because the claimed inter-variate gains are attributed to the prototype interaction. Please provide evidence that RD separates normal from anomalous timestamps, for example by reporting RD distributions on normal versus anomalous test segments or by replacing Eq. (17) with TD-only and with random-prototype versions.","section":"§2.6–§2.7, Eq. (17)"},{"comment":"The text states that MtsCID 'secures the highest F1 and AF-F1 scores in six out of seven datasets,' but Table 3 does not support this reading. On SWaT, MtsCID's AF-F1 is 57.01, well below STEN's 70.98 and DCdetector's 69.75; on SMD, STEN's VUS-ROC is 91.29 versus MtsCID's 84.22; on SMAP, STEN's VUS-PR and VUS-ROC are higher than MtsCID's. The summary of results should be qualified to name the specific metric-dataset combinations where MtsCID is best, second-best, or not competitive, rather than claiming blanket superiority.","section":"§4.1, Table 3"},{"comment":"The paper says 'Each method was executed five times for each dataset, and the resulting values were averaged,' but the note under Table 2 states that results for iForest, DeepSVDD, DAGMM, THOC, and InterFusion were adopted from [30] rather than re-run. No standard deviations are reported for any method, so the statement that MtsCID outperforms all nine baselines is not backed by comparable experimental evidence. Please either re-run the five copied baselines under the same protocol or clearly separate reproduced results from cited results, and report variance across the five runs.","section":"§4.1, Table 2 note"},{"comment":"The reported F1 and AF-F1 values use the best threshold chosen on the test set, which can inflate performance and makes cross-method comparisons optimistic. Since the paper also reports VUS-ROC and VUS-PR, which are threshold-free, please state explicitly that the headline F1 comparison is an upper-bound estimate, or select thresholds on the validation split and report those F1 values as the primary comparison.","section":"§3.2, §4.1"},{"comment":"Equation (17) is not dimensionally well-defined as written: AScore(X) is declared to be in R^L, but TD(X_t,:, \\hat X_t,:) is a distance between C-dimensional vectors and RD(O_t,:, M_::) is also indexed per time step. It is unclear whether TD is first aggregated over variates, whether RD is a vector of length C or L, and how the softmax over RD is aligned with the element-wise multiplication. Please provide explicit definitions of TD and RD as functions with stated output shapes, and write Eq. (17) with the correct indices.","section":"§2.7, Eq. (17)"}],"minor_comments":[{"comment":"The first contribution contains a typo: 'an novel scheme' should be 'a novel scheme.'","section":"§1, Contributions"},{"comment":"The figure label 't-AutoEcoder' is misspelled; it should be 't-AutoEncoder.'","section":"Figure 1"},{"comment":"The notation w_ti is introduced with indices t and i, but the summation in the denominator runs over j=1..L; please clarify that j indexes the L rows of M and align the index names consistently.","section":"§2.5, Eq. (13)"},{"comment":"The discussion of why MtsCID works is qualitative and does not reference any quantitative evidence beyond the ablation table; consider adding a short analysis of which component drives the gain on each dataset.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hey [Name], quick take on arXiv:2501.16364.\n\nMtsCID is a solid entry in the crowded MTS anomaly detection space. The architecture is clearly described: a dual-branch setup with frequency-domain transformers, multi-scale patch attention on attention maps, and a fixed sinusoidal prototype memory. What's genuinely new is the specific combination and the idea of using coarse-grained attention maps for intra-variate dependencies. The empirical work is honest: seven benchmarks, multiple metrics, ablations, sensitivity analysis, and the code is public. The F1 gains are modest on most datasets, which is realistic for this mature field; the big jump on GECCO is notable but on a niche dataset.\n\nThe soft spots are real. The most serious is the anomaly score in Eq. 17. The Relationship Deviation term RD is never minimized during training. The lower branch is trained only with an entropy loss on the softmax attention weights over the fixed sinusoidal prototypes. That loss encourages the attention distribution to be peaky, but it does not make RD small, calibrated, or even meaningful. The softmax over RD in the anomaly score normalizes the scale away, so the weighting reduces to a relative ranking of RD values across time steps. If all normal time steps have similar RD, the term does nothing beyond TD. The ablation that removes the lower branch conflates the RD scoring mechanism with the extra network capacity. The paper should show RD distributions for normal vs. anomalous timestamps, or an ablation that removes RD from the score while keeping the lower branch. This is not a fatal flaw for the whole paper, but it does mean the claimed benefit of sinusoidal prototypes is unsupported as presented.\n\nOther issues: no standard deviations over the five runs, which makes it hard to judge whether the one-point F1 leads over MEMTO are meaningful. Best-F1 on the test set is standard practice in this literature but worth noting. Five baselines are taken from a prior paper rather than re-run; less ideal but acceptable given code availability. Also, there's a dimension inconsistency for M in Sec. 2.5 (R^{C×L} vs R^{L×C})—minor but easy to fix.\n\nOverall, this is a solid empirical paper that deserves a serious referee. The architecture is original enough, the code is there, and the experiments are thorough. The RD concern should be addressed in revision, but it doesn't invalidate the whole approach. I'd send it to review and ask for the RD analysis plus error bars.\n\nRecommendation: accept for peer review with major revision expected.","headline":"A competent and honest empirical paper with public code, but the claimed mechanism for the sinusoidal prototype memory is not actually trained to make the relationship-deviation term meaningful, so the inter-variate gains are not fully supported.","tokens_in":14910,"tokens_out":2779,"would_cite":false,"duration_ms":28604,"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":"MtsCID's coarse-grained dual-network design, using multi-scale patch attention and fixed sinusoidal prototypes, beats nine baselines on six of seven anomaly detection datasets.","keywords":["multivariate time series anomaly detection","semi-supervised learning","coarse-grained dependencies","multi-scale patch attention","sinusoidal prototypes","frequency domain processing","AIOps"],"falsifier":"Take a synthetic multivariate process with no periodicity, where normality is defined by a linear subspace such as $x_1 + x_2 + \\dots + x_C = 0$ and anomalies violate that constraint; train MtsCID on normal data and check whether the relationship-deviation term (RD in Equation 17) rises at anomaly timestamps. If RD does not separate the classes, the sinusoidal prototype memory is insufficient for non-periodic inter-variate structure. A complementary test is to repeat the GECCO and SWAN experiments with a learned prototype memory in place of the fixed sinusoids; if the learned memory matches or beats MtsCID there, the fixed basis is the limiting factor.","tokens_in":13978,"feed_emoji":"📈","tokens_out":7123,"duration_ms":69670,"temperature":0.7,"pith_summary":"The paper sets out to show that multivariate time series anomaly detection is better served by coarse-grained dependency learning than by fine-grained step-level modeling. MtsCID pairs a temporal autoencoder branch, which learns from attention maps of multi-scale patches in the frequency domain, with an inter-variate branch that combines convolution, a frequency-domain transformer along the variate dimension, and a fixed sinusoidal prototype memory. Training uses a reconstruction loss plus an entropy loss on prototype weights, and anomalies are scored from temporal deviation and relationship deviation. On seven public datasets, the reported F1 is the best in six and second-best in one, against nine state-of-the-art baselines. A reader should care because reliable anomaly detection underpins failure management in web operations and critical infrastructure.","feed_headline":"Coarse-grained design wins six of seven anomaly benchmarks","feed_subtitle":"MtsCID captures dependencies at patch level, not step level, and beats nine state-of-the-art methods.","key_machinery":"The load-bearing machinery is the pair of coarse-grained encoders plus the sinusoidal prototype memory. The t-AutoEncoder (temporal autoencoder) computes the discrete Fourier transform, projects real and imaginary parts through fc-Linear and fc-Transformer in frequency space, returns to the time domain, and forms attention maps $A_{p_i} = \\mathrm{Softmax}(Z_{p_i}Z_{p_i}^T / \\sqrt{p_i})M_{p_i}$ over multi-scale patches; these maps are the features used for temporal dependency learning. The i-Encoder (inter-variate encoder) applies a 1D convolution over the time axis, then a frequency transformer whose self-attention runs along the variate dimension, yielding representations $O$. The p-i Module (sinusoidal prototypes interaction) multiplies $O$ with the fixed matrix $M_{i,j} = \\cos(2\\pi i j / L)$, which encodes normal inter-variate combinations as a limited set of periodic patterns; a softmax over prototype positions gives weights whose entropy is minimized during training. These components together convert the problem from fine-grained step reconstruction into coarse-grained pattern matching against a fixed periodic memory.","core_discovery":"The paper's central claim is that coarse-grained intra-variate and inter-variate dependencies are the missing ingredient in semi-supervised multivariate time series anomaly detection. Step-level methods reconstruct or score each timestamp, which the authors argue is too fine-grained to capture salient normal patterns. MtsCID instead learns temporal dependencies from attention maps of multi-scale patches applied to frequency-domain representations, and learns inter-variate relationships by passing a 1D-convolved time-domain representation through a frequency transformer along the variate axis, then interacting it with a fixed matrix of sinusoidal prototypes, $M_{i,j} = \\cos(2\\pi i j / L)$. The anomaly score multiplies temporal reconstruction deviation by a softmax-weighted relationship deviation, measuring how far each time step's inter-variate state sits from the nearest prototype. The paper reports that this design outperforms nine baseline methods on seven widely used datasets, including a 42.12 percent F1 improvement over the second-best method on the GECCO dataset.","pith_inferences":["We infer that the fixed sinusoidal prototype basis is a strong inductive bias; a dataset whose normal cross-variate relationships are non-periodic, such as those governed by arbitrary graph structure, would likely require learned or data-adaptive prototypes, a regime the paper does not test.","The coarse-graining principle, using attention-map features rather than step-level features, could transfer to other reconstruction-based and contrastive time series anomaly detectors, not just this architecture.","The frequency-domain alignment argument suggests a natural application to unsynchronized multi-sensor streams, where aligning variates in frequency space could compensate for loose time synchronization; the paper asserts this benefit but does not measure it directly.","SWaT, where MtsCID is second-best, is a continuous physical process, so the sinusoidal prototype assumption may be less decisive there; that dataset offers a natural test case for prototype design."],"forward_implications":["If correct, anomaly detectors for cloud and web operations can rely on segment-level and cross-variate pattern deviations instead of per-step reconstruction, which should be more tolerant of noise and sensor misalignment.","The fixed sinusoidal prototype memory removes the need for learned memory updates and two-phase training, simplifying deployment while retaining prototype-based inter-variate modeling.","The method's strong gain on GECCO indicates coarse-grained inter-variate modeling helps when anomalies are rare and heterogeneous, not just when they follow clean periodic patterns.","Because the two branches are independent, the architecture can be extended by upgrading either branch, for example with richer patch hierarchies or learned prototype bases, without redesigning the other."],"supporting_citations":[{"why":"MEMTO memory-guided transformer; provides the prototype interaction idea, entropy loss, and subsequence-generation protocol that MtsCID builds on.","marker":"[30]"},{"why":"DCdetector; supplies patch-based attention contrastive learning baseline and the GECCO/SWAN dataset curation followed here.","marker":"[37]"},{"why":"STEN; gives a self-supervised spatiotemporal baseline and the best-F1 evaluation protocol used for comparisons.","marker":"[5]"},{"why":"AT (Anomaly Transformer); the key temporal association baseline that MtsCID must beat on most datasets.","marker":"[36]"},{"why":"InterFusion; motivates capturing inter-variate relationships and provides a hierarchical VAE baseline.","marker":"[15]"},{"why":"Hundman et al.; contributes the SMAP and MSL spacecraft telemetry datasets used in evaluation.","marker":"[11]"},{"why":"Su et al.; contributes the Server Machine Dataset (SMD) used in evaluation.","marker":"[31]"},{"why":"Mathur and Tippenhauer; contributes the SWaT water-treatment dataset used in evaluation.","marker":"[20]"},{"why":"Moritz et al.; contributes the GECCO drinking-water dataset used in evaluation.","marker":"[21]"}],"fun_headline_variants":["Coarse-grained temporal and inter-variate learning tops anomaly tests","MtsCID uses coarse-grained dependencies to outscore nine baselines","Patch-level dependencies win six of seven anomaly detection benchmarks","Coarse-grained view of time and variables yields stronger anomaly scores","MtsCID captures coarse patterns to outperform fine-grained baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed sinusoidal prototype matrix, $M_{i,j} = \\cos(2\\pi i j / L)$, contains a sufficient set of normal inter-variate combinations for every dataset; if true normal relationships do not live near these periodic prototype combinations, the relationship-deviation term in the anomaly score cannot separate normal from anomalous states, and the reported inter-variate gains would vanish.","fun_headline_variants_meta":{"raw":{"variants":["Coarse-grained temporal and inter-variate learning tops anomaly tests","MtsCID uses coarse-grained dependencies to outscore nine baselines","Patch-level dependencies win six of seven anomaly detection benchmarks","Coarse-grained view of time and variables yields stronger anomaly scores","MtsCID captures coarse patterns to outperform fine-grained baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000684,"raw_usage":{"total_tokens":3129,"prompt_tokens":994,"completion_tokens":2135,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":610,"tokens_out":2135,"duration_ms":17099,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:55:24.813489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic multivariate process with no periodicity, where normality is defined by a linear subspace such as $x_1 + x_2 + \\dots + x_C = 0$ and anomalies violate that constraint; train MtsCID on normal data and check whether the relationship-deviation term (RD in Equation 17) rises at anomaly timestamps. If RD does not separate the classes, the sinusoidal prototype memory is insufficient for non-periodic inter-variate structure. A complementary test is to repeat the GECCO and SWAN experiments with a learned prototype memory in place of the fixed sinusoids; if the learned memory matches or beats MtsCID there, the fixed basis is the limiting factor.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MEMTO memory-guided transformer; provides the prototype interaction idea, entropy loss, and subsequence-generation protocol that MtsCID builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"STEN; gives a self-supervised spatiotemporal baseline and the best-F1 evaluation protocol used for comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"InterFusion; motivates capturing inter-variate relationships and provides a hierarchical VAE baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mathur and Tippenhauer; contributes the SWaT water-treatment dataset used in evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Moritz et al.; contributes the GECCO drinking-water dataset used in evaluation."}],"review_version":1}