acceptodds
Under review as a conference paper at ICLR 2027

CHIME: CONTEXTUAL HIERARCHICAL INCONSISTENCY MODELING FOR MULTIVARIATE TIME SERIES ANOMALY DETECTION

Abstract

Reconstruction-based anomaly detection assumes that anomalous observations are harder to reconstruct than normal ones. In multivariate time series, this assumption can fail when anomalies disrupt temporal or inter-variable relationships while remaining accurately reconstructable. This paper presents CHIME, a framework for Contextual Hierarchical Inconsistency Modeling that incorporates historical context into the assessment of cross-scale structural discrepancies. Preceding windows are encoded into context tokens that guide bidirectional cross-attention among fine-, middle-, and coarse-scale representations. Coarse-to-fine attention relates local patterns to broader trends, while fine-to-coarse attention updates coarse representations with local evidence. A variable-relation branch captures dependencies at each resolution. The anomaly score combines reconstruction error with two forms of hierarchical inconsistency: discrepancies between aligned temporal representations and discrepancies between variable-relation representations across scales. Experiments on ten public benchmarks show that CHIME achieves a mean point-adjusted F1 score of 95.54%, ranking first on eight datasets among the compared methods.

Primary area:
learning on time series and dynamical systems
Keywords:
Anomaly Detection, Representation Learning, Multivariate Time Series, Context-Aware Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kJ2xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper presents CHIME (Contextual Hierarchical Inconsistency Modeling) for multivariate time-series anomaly detection, addressing anomalies that disrupt temporal or inter-variable relationships while remaining accurately reconstructable.

  • CHIME encodes preceding windows into context tokens that condition multi-resolution temporal representations, which then interact through coarse-to-fine and fine-to-coarse cross-attention. A scale-specific variable-relation branch captures inter-variable dependencies at each resolution. The anomaly score combines reconstruction discrepancy with temporal and variable-relation inconsistency across scales.

  • On ten public benchmarks, CHIME achieves a mean point-adjusted F1 score of 95.54% and ranks first on eight datasets among the compared methods. Ablations on four datasets show that removing historical context or bidirectional scale interaction reduces average F1, whereas adding the temporal and variable-relation inconsistency scores to reconstruction improves it, with the full combination performing best. Complementary AF-F1, VUS-PR, and AUC-ROC results are also reported on five representative datasets.

Critical strengths

1. Hierarchical inconsistency as complementary anomaly evidence

  • Rather than using different scales or variable relations mainly as additional representations, CHIME explicitly measures whether their structures remain consistent, through temporal inconsistency across aligned representations and variable-relation inconsistency across scales. The score-component ablation shows that each inconsistency term improves over reconstruction alone, while their combination performs best on all four ablation datasets.

2. Broad empirical evaluation

  • CHIME is evaluated on ten public benchmarks against ten representative baselines covering reconstruction-, association-, multi-scale-, and patch-based methods. It ranks among the top two methods on every benchmark in terms of point-adjusted F1. Module ablations further examine the contributions of historical context and bidirectional scale interaction .
Critical weaknesses

1. Limited direct validation of detecting reconstructable structural anomalies

  • CHIME is motivated by anomalies that disrupt temporal or inter-variable relationships while remaining accurately reconstructable (Abstract; §1). However, the score-component ablation (Table 4) demonstrates improvements in overall F1 without directly establishing whether the inconsistency scores detect anomalies with low reconstruction error. It also remains unclear how context-dependent inconsistency distinguishes such anomalies from legitimate operating-state transitions. The authors identify robustness to changing operating states as future work (§5).

2. Incomplete component attribution and evaluation robustness

  • The module ablation (Table 3) removes historical context and bidirectional scale interaction as a whole. It therefore leaves unclear whether both attention directions are needed and whether the context gain comes from historical information rather than additional model capacity.

  • The main comparison emphasizes point-adjusted F1, while the main text does not specify how the threshold τ\tau is selected (§3.6). Event- and range-aware metrics are reported on five datasets without Anomaly Transformer or DCdetector, the only baselines that outperform CHIME on any dataset in Table 1. On TAO, CHIME achieves substantially lower VUS-PR than CrossAD despite higher point-adjusted F1 (74.01% vs. 95.68%; Table 2). Full-model VUS-PR also differs between Tables 2 and 3 on SWaT and TAO, without a stated explanation.

Questions for authors

Q1. Detection of reconstructable structural anomalies (Weakness 1)

  • Among anomalous positions whose reconstruction score SrecS_{\mathrm{rec}} remains within the normal range, what fraction can be detected by incorporating StempS_{\mathrm{temp}} and SrelS_{\mathrm{rel}}? Could the authors also report false-positive rates on normal segments following operating-state changes, with and without historical context?

Q2. Contributions of bidirectional interaction and historical context (Weakness 2)

  • Could the authors compare coarse-to-fine, fine-to-coarse, and bidirectional cross-scale interaction to clarify the contribution of each direction (Eqs. 4–5; Table 3)?

  • Does the benefit of historical context come from preceding operating-state information rather than additional model capacity, for example under shuffled-context or capacity-matched comparisons?

Q3. Evaluation protocol and robustness (Weakness 2)

  • How is the threshold τ\tau selected (§3.6), and does the selection use test labels? Could the authors report event- and range-aware metrics for Anomaly Transformer and DCdetector, at least on TAO and Exathlon, given the gap between point-adjusted F1 and VUS-PR on TAO (Table 2)?

  • Could the authors clarify why the full-model VUS-PR results differ between Tables 2 and 3 on SWaT and TAO?

5 comments
AuthorsOpenReview ↗

Response to Q1: Low-Error Detection and False Positives

Thank you for requesting direct point-level recovery statistics and false-positive rates around normal operating-state transitions. We performed a NIPS-TS-Water diagnostic using the existing context-enabled checkpoint, without retraining.

We define low reconstruction error using the final decoder output's raw MSE RtR_t, with threshold τR=0.07747\tau_R=0.07747 (95th percentile of normal-validation RtR_t). Among 730 ground-truth anomalous points, 305 (41.78%) satisfy Rt≤τRR_t\leq\tau_R. Note that this diagnostic quantity is different from Eq. (16)'s SrecS_{\mathrm{rec}}, which aggregates reconstruction errors from multiple decoding scales. Accordingly, a point with low raw MSE may still be detected by reconstruction-only scoring.

For each score, the detection threshold is independently fixed at its 95th percentile on normal-validation data, with neither point adjustment nor test-label-based threshold tuning:

NIPS-TS-Water diagnosticCount / fraction
Ground-truth anomalous points730
Low-raw-MSE anomalous points305/730 (41.78%)
Detected by SrecS_{\mathrm{rec}} alone247/305 (80.98%)
Detected by Full CHIME248/305 (81.31%)
Additional detection by Full1/305 (0.33 percentage points)

Thus, the final combined score yields only a very small additional recovery under this checkpoint and thresholding protocol. We will not present this as evidence that CHIME substantially recovers all reconstructable anomalies. The separate 20/22 event-level union of temporal and relation component responses reported in R2-Q2 is not the final model's event detection rate and cannot replace these point-level statistics.

Regarding false positives after legitimate operating-state transitions, we agree that such changes may themselves produce cross-scale discrepancies. The available annotations/logs do not reliably identify normal state-transition intervals, so we cannot report a defensible transition-specific false-positive rate for Full versus w/o Context. We will state this limitation explicitly rather than substitute overall normal-point false-positive rates or claim that history prevents these false alarms.

AuthorsOpenReview ↗

Response to Q2: Bidirectional Interaction and Context Ablations

Thank you for proposing targeted ablations. We conducted separate comparisons of attention direction and historical-information content.

  1. Directional interaction. The PA-F1 results (mean ±\pm standard deviation) are:
DatasetCoarse-to-fineFine-to-coarseBidirectional
NIPS-TS-Water0.891 ±\pm 0.0120.898 ±\pm 0.0100.9102 ±\pm 0.008
SMD0.934 ±\pm 0.0050.936 ±\pm 0.0040.9423 ±\pm 0.004
SWaT0.978 ±\pm 0.0020.979 ±\pm 0.0020.9816 ±\pm 0.001
TAO0.913 ±\pm 0.0120.921 ±\pm 0.0100.9443 ±\pm 0.007
Mean0.92900.93350.9446

Bidirectional interaction improves mean PA-F1 over coarse-to-fine alone by 1.56 percentage points and over fine-to-coarse alone by 1.11 points. The former brings coarse-scale trends to fine representations; the latter feeds fine-scale local responses back to coarse representations. The benefit is dataset dependent, especially small on SWaT, and we do not infer statistical significance from means alone.

  1. Historical information versus extra capacity. Since removing context changes both inputs and parameters, we added E3 (context-related capacity retained, no informative history) and E2 (history mismatched to the current window):
DatasetNo contextE3E2Correct history
NIPS-TS-Water0.87960.88300.88930.9102
SMD0.93910.93950.94050.9423
SWaT0.97920.97930.98010.9816
TAO0.88920.89500.90500.9443
Mean PA-F10.92180.92420.92870.9446

Correctly aligned history outperforms E3 and E2 by 2.04 and 1.59 percentage points on average. This supports the interpretation that preceding observations offer sample-specific information beyond merely retaining model capacity. The differences are much smaller on SMD and SWaT, and these controlled comparisons should not be interpreted as proof excluding all training/capacity effects. We will document the E2/E3 construction and training settings in the revision.

AuthorsOpenReview ↗

Response to Q3: Evaluation Protocol and Metric Corrections

Thank you for pointing out missing protocol details and inconsistencies. We audited the evaluation code and result tables.

  1. Thresholding and label use. For PA-F1, we apply ground-truth-interval point adjustment and then choose the threshold maximizing F1 on adjusted test scores using test labels. This protocol is applied uniformly to all baselines, with a separately optimized threshold for each method's score scale. Thus, PA-F1 is explicitly a test-label-assisted retrospective metric. Reported AF-F1 uses the 99th percentile of raw test scores without PA; this percentile uses the test-score distribution but not test labels for its selection. VUS-PR, VUS-ROC, and AUC-ROC are calculated from continuous anomaly scores without a single fixed threshold.

  2. Strong-baseline comparison. The following results are from the complete evaluation tables (percent):

DatasetMethodPA-F1AF-F1VUS-PR
TAOCHIME94.4375.1874.01
TAOAnomaly Transformer96.4578.2492.44
TAODCdetector55.8864.3340.68
TAOCrossAD92.6591.5195.68
ExathlonCHIME98.0968.2695.16
ExathlonAnomaly Transformer95.8432.1093.31
ExathlonDCdetector99.4671.2799.48
ExathlonCrossAD58.8415.6642.54

These comparisons show that PA-F1 ranking is not preserved under event- or range-aware metrics. On TAO, for example, CHIME's PA-F1 is higher than CrossAD's, but its AF-F1 and VUS-PR are substantially lower. We will highlight these differences rather than assert superiority across all metrics.

  1. Table 2/3 discrepancy. The inconsistent Full-model VUS-PR values for SWaT and TAO are transcription errors, not differences in metric definition or evaluation protocol. The correct values are 96.85% for SWaT and 74.01% for TAO (as in Table 2); Table 3's 97.88% and 73.24% should be corrected. We apologize for these mistakes and will verify the remaining table entries in the revision.
Reviewer_X9iZOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CHIME (Contextual Hierarchical Inconsistency Modeling) for multivariate time-series anomaly detection. The authors argue that, in multivariate time series, anomalies can disrupt temporal or inter-variable relationships while still being accurately reconstructed, so reconstruction error alone may fail to detect them.

  • Historical context modeling: Preceding windows are encoded into context tokens, which are injected into multi-resolution tokens through a gated residual.
  • Hierarchical cross-scale modeling:
    • Multi-resolution tokens are constructed with several pooling factors.
    • Bidirectional (coarse-to-fine and fine-to-coarse) cross-attention is applied between adjacent resolutions.
    • A progressive decoder reconstructs the input across resolutions.
  • Hierarchical inconsistency scoring: The final anomaly score is a weighted sum of three MAD-normalized terms:
    • the reconstruction error;
    • the discrepancies between latent representations and decoder outputs at adjacent resolutions (temporal inconsistency);
    • the discrepancies between scale-specific variable-relation matrices (variable-relation inconsistency).

On ten public benchmarks, the paper reports a mean point-adjusted F1 of 95.54% and claims first place on eight datasets.

Critical strengths

S1 (Quality): The method is evaluated on a diverse set of benchmarks.

  • Experiments are conducted on ten public benchmarks with different characteristics, including spacecraft telemetry, server monitoring, industrial control, and environmental sensing.

S2 (Quality): Some metrics beyond point adjustment are reported.

  • Range- and event-aware metrics such as VUS-PR, AF-F1, and AUC-ROC are reported on representative datasets.

S3 (Quality): The contribution of each score component is shown.

  • The score-component ablation shows the effect of adding temporal and variable-relation inconsistency to the reconstruction error.

S4 (Clarity): Limitations and future directions are stated.

  • The Conclusion identifies remaining challenges, such as the conditions under which the inconsistency signals are beneficial and robustness to contaminated context.
Critical weaknesses

W1 (Originality): The method is close to a combination of existing approaches, and its differences from the closest prior work are unclear.

  • The proposed method is close to a combination of existing components:
    • reconstruction
    • association and inter-variable relation modeling
    • multi-scale representations
    • historical-window context
  • The authors themselves describe CrossAD as explicitly modeling cross-scale associations and incorporating cross-window information. The novelty of CHIME therefore cannot lie simply in "using both scales and preceding windows."
  • However, the paper does not show how CHIME's cross-scale inconsistency scores differ, formally or in practice, from CrossAD's cross-scale associations.

W2 (Significance): The core motivation and problem statements are not sufficiently substantiated.

  • The authors argue that anomalies can be accurately reconstructed while disrupting relationships. However, a reconstruction model that has learned inter-variable dependencies well may instead produce larger errors for inputs whose relationships are broken. The statement that "reconstruction error mainly describes numerical deviations" should therefore be presented as the specific failure case the authors target, not as a general property of reconstruction methods.
  • To substantiate the motivation, the paper should directly show that the proposed inconsistency scores detect anomalies whose reconstruction errors are low.
  • The claim that "most existing methods use different scales or variable relations mainly as additional representations, rather than explicitly measuring whether their structures remain consistent" also lacks support. For example, Anomaly Transformer directly uses association discrepancy as its score, and CrossAD explicitly models cross-scale associations.
  • The paper also does not explain, through concrete examples or figures, why temporal inconsistency across resolutions and variable-relation inconsistency across scales are needed.

W3 (Soundness): The validity of the variable-relation inconsistency score is unclear.

  • SrelS_{\text{rel}} does not directly measure changes in actual variable dependencies.
    • What Eq. (11) directly captures is only that the learned attention matrices differ across scales.
    • Interpreting this as a change in actual sensor dependencies requires either validation on data with known relational changes or evidence that distinguishes normal scale differences from anomaly-induced ones.
  • The authors state that the relation matrices at different resolutions need not be identical.
    • "This score measures changes in cross-resolution variable dependencies without requiring relation graphs at different resolutions to be identical."
    • However, Lrel\mathcal L_{\text{rel}}, trained on normal data, continually reduces the difference between the two matrices, and this term is minimized when they are identical.
    • If relations differ systematically across scales even under normal operation, the paper should explain why this difference should be reduced.
  • Eq. (10) projects a single scalar at each position into queries and keys.
    • If this projection is affine, each row reduces to a softmax over the scalar values at the current position.
    • In that case, the matrix encodes neither variable identity nor learned pairwise dependency structure, which makes it difficult to interpret as a "variable-relation graph."

W4 (Soundness): The design of the inconsistency scores and the context module is insufficiently justified.

  • CHIME explicitly computes differences between adjacent scales. However, these are heuristic distances in latent space rather than tests of structural compatibility.
  • These distances are computed on representations that have already been mixed across adjacent scales through bidirectional cross-attention.
    • The fusion may therefore reduce anomaly-induced differences between scales.
    • Moreover, Ltemp\mathcal L_{\text{temp}} explicitly reduces these differences on normal data.
  • The normal patterns of preceding windows are already reflected during training. The paper does not explain what additional information the separate context module provides beyond this, and it does not compare against the simple alternative of using a longer window.
  • Eq. (1) applies a shared function fwf_w to each of the KK preceding windows and takes a weighted sum over windows.
    • Because no information about window order is used, permuting the preceding windows yields the same result.
    • This is inconsistent with the motivation of comparing against the immediately preceding system states.

W5 (Quality): The evaluation protocol does not sufficiently support the main results.

  • The headline result in the abstract and the comparison across all ten datasets (Table 1) rely solely on point-adjusted F1. Point-wise F1 is not reported.
  • Point adjustment is known to yield high F1 even for nearly random scores (Kim et al., AAAI 2022), as has been observed for methods such as Anomaly Transformer. PA-F1 alone is therefore not a reliable indicator of detection performance.
  • Table 2 includes only a subset of the baselines.
    • In particular, DCdetector and Anomaly Transformer, the strongest competitors in terms of PA-F1 in Table 1, are missing.
    • All baselines should be compared fairly under the same metrics.
  • Unlike Table 1, Table 2 does not mark the second-best results (underline).

W6 (Quality): The ablations and statistical support are insufficient.

  • In the module ablation (Table 3), the claim that removing a module degrades performance holds consistently only for F1.
  • Table 3 does not mark the best results for metrics other than F1, and Table 4 reports only F1.
  • There is no single-scale ablation that removes the multi-scale structure itself.
  • Figure 2 presents small differences only as plots; a table would be more appropriate.
  • All results come from single runs without standard deviations, so the small differences observed in the ablations cannot be distinguished from seed variation.

W7 (Reproducibility): The detection procedure is underspecified.

  • The procedure for choosing the threshold τ\tau is not described.
  • There is no rationale for the score-fusion weights (0.8, 0.1, 0.1), and the "fine-dominant" setting in Figure 2(c) is not defined.

W8 (Clarity): The figures and notation are unclear.

  • Figure 1 is hard to grasp at a glance, and the text and equations inside it are not neatly organized. The progressive decoder is also not depicted.
  • Several symbols are not defined when they first appear (e.g., dhd_h, fwf_w, WhW_h, ηctx\eta_{\text{ctx}}, fgf_g, drd_r, pp, UjU_j).
Questions for authors

Q1 (W1): What is the methodological novelty of CHIME, and how does it differ from CrossAD, the closest prior work?

  • CHIME combines reconstruction, inter-variable relation modeling, multi-scale representations, and historical-window context. Beyond this combination, what do you consider its core methodological contribution?
  • From a research perspective, how do CHIME's cross-scale inconsistency scores differ from CrossAD's cross-scale associations?

Q2 (W2): Can the core motivation, namely anomalies that are accurately reconstructed, be demonstrated directly?

  • Can you quantify how many anomalous segments with low reconstruction error exist in each dataset?
  • Can you show, through case studies or quantitative analysis, that the temporal and variable-relation inconsistency scores detect anomalies in such segments?
  • Is the statement that "reconstruction error mainly describes numerical deviations" a general property of reconstruction methods, or a failure case that arises only under specific conditions? Is there evidence that this problem also occurs for reconstruction models that have learned inter-variable dependencies well?
  • Methods such as Anomaly Transformer (association discrepancy) and CrossAD (cross-scale associations) already measure structural differences directly. What is the basis for the claim that most existing methods do not measure structural consistency?
  • Can you illustrate, with concrete examples or figures, situations in which temporal inconsistency across resolutions and variable-relation inconsistency across scales are needed?

Q3 (W3): Does the variable-relation inconsistency score reflect actual changes in variable dependencies?

  • What exactly is the form of the projection of each scalar value into queries and keys in Eq. (10)?
    • If this projection is affine, the relation matrix reduces to a softmax over the current values.
    • Can it still represent variable identity and pairwise dependency structure in that case?
  • How do you distinguish relational differences across scales that occur in normal data from those caused by anomalies?
  • The authors state that the relation matrices need not be identical.
    • Why, then, is the model trained with Lrel\mathcal L_{\text{rel}} to reduce their differences on normal data?
    • How does the detection performance of SrelS_{\text{rel}} change when αrel=0\alpha_{\text{rel}}=0?

Q4 (W4): What is the rationale for the design of the inconsistency scores and the context module?

  • What is the basis for assuming that distances between scales computed in latent space reflect structural compatibility?
  • The inconsistency scores are computed on representations that have already been mixed across adjacent scales through bidirectional cross-attention. How do the results change if inconsistency is computed on representations before cross-scale fusion?
  • Ltemp\mathcal L_{\text{temp}} is trained to reduce scale differences on normal data. Does it also reduce anomaly-induced differences?
  • The normal patterns of preceding windows are already reflected during training. What additional information does the separate context module provide beyond this?
  • Eq. (1) does not distinguish the order of preceding windows.
    • Is this an intended design choice?
    • How do the results change if window-order information (e.g., positional encoding) is added?

Q5 (W5, W6): Do the results hold under evaluation that does not rely on point adjustment, and are they statistically stable?

  • It would be helpful to report point-wise F1 together with metrics such as VUS-PR and AF-F1 for all ten datasets.
  • Could all baselines, including DCdetector and Anomaly Transformer, be compared under the metrics of Table 2, with the best and second-best results marked as in Table 1?
  • What are the means and standard deviations over multiple seeds? In particular, do the small differences in Tables 3 and 4 exceed seed variation?
  • How should the results of the module ablation (Table 3) be interpreted for metrics other than F1? Please also report the score-component ablation in Table 4 for these metrics.
  • How does a single-scale variant, with the multi-scale structure removed, perform?
  • Providing the parameter-sensitivity results in Figure 2 as a table with numerical values would make the comparison clearer.
9 comments
AuthorsOpenReview ↗

Response to Q1: Methodological Novelty

Thank you for requesting a clearer methodological contribution. We agree that reconstruction, multi-scale representations, inter-variable modeling, and historical windows are individually established techniques; their simple combination is not our novelty claim.

CHIME's central design is to use the normal cross-scale responses learned under historical-context conditioning as a reference and then explicitly score residual discrepancies between adjacent resolutions. The temporal score compares both aligned latent representations and progressive decoder outputs. The relation score compares scale-specific attention-based relation proxies. These two scores enter the final decision alongside multi-scale reconstruction error, rather than serving only to enrich features used by the decoder.

The distinction from CrossAD is therefore not the existence of cross-scale associations. CrossAD already leverages them in its representation/reconstruction process. CHIME instead makes post-interaction temporal and relation-proxy inconsistencies explicit auxiliary detection signals. Historical conditioning gives these comparisons sample-dependent preceding-state information.

We also recognize that earlier approaches, including Anomaly Transformer, use association discrepancies as anomaly evidence. We do not claim to originate this general principle or to recover true physical dependency changes from attention scores. Our contribution is the specific context-conditioned hierarchical scoring mechanism and its integration with reconstruction. We will revise the novelty claims and strengthen the comparison to CrossAD accordingly.

AuthorsOpenReview ↗

Response to Q2: Low-Reconstruction-Error Anomalies

Thank you for asking for direct evidence of anomalies that remain well reconstructed. We agree that the statement that reconstruction errors mainly describe numerical deviations was overly broad: a reconstruction model that learns temporal or inter-variable dependencies can also assign large errors to dependency violations. Our motivation concerns failure cases under particular models and data conditions, not a universal limitation of reconstruction.

  1. Event-level component diagnosis. We examined four datasets and counted anomalous events with low reconstruction error, together with events exhibiting a response from each auxiliary score:
DatasetAnomaly eventsLow-error eventsTemporalRelationUnion
NIPS-TS-Water2243/42/44/4
SMD6085/85/87/8
SWaT3653/53/54/5
TAO3053/53/55/5
Total1482214/2213/2220/22

The 22 low-error events comprise 14.86% of the 148 anomalous events. The two scores respond to partially different events, suggesting component-level complementarity. Critically, the 20/22 union is NOT the detection rate of the final fused CHIME score and does not imply recovery of all anomalous points. For example, in NIPS-TS-Water interval [5248, 5271), the temporal score responded despite low raw reconstruction error, but the fused detector missed the event.

  1. Final-score recovery at the point level. In a separate NIPS-TS-Water diagnostic using an existing checkpoint, we defined low raw reconstruction MSE by the 95th percentile of normal-validation MSE, τR=0.07747\tau_R=0.07747. Of 730 anomalous points, 305 (41.78%) met Rt≤τRR_t\leq\tau_R. Among those 305 points, reconstruction-only detected 247 (80.98%) and full CHIME detected 248 (81.31%): just one additional point (0.33 percentage points). Both configurations used their own 95th-percentile normal-validation score thresholds, without point adjustment or test-label-based threshold optimization. Here RtR_t denotes the final output's raw MSE, whereas SrecS_{\mathrm{rec}} aggregates multiple decoder-scale errors; low RtR_t does not imply low SrecS_{\mathrm{rec}}.

Thus, the component responses do not necessarily translate into additional final detections. The current evidence does not support a claim of substantial recovery of reconstructable anomalies.

  1. Scope and prior work. Benchmark anomaly labels do not identify ground-truth changes in physical or causal dependencies, so we cannot establish that every low-error event is a genuine dependency-change anomaly. Anomaly Transformer already exploits association discrepancy, and CrossAD models cross-scale association. CHIME does not claim to invent structural discrepancy scoring; it examines context-conditioned temporal and relation-proxy discrepancies as auxiliary evidence. We will narrow the motivation and clearly distinguish event-level component diagnostics from point-level final detections.
AuthorsOpenReview ↗

Response to Q3: Variable-Relation Inconsistency

We agree that the original description overinterpreted the relation matrices. We clarify the parameterization, interpretation, and effect of the auxiliary loss.

  1. Scalar Q/K projections. In Eqs. (9)–(10), a scale-specific scalar signal vn=Vt(s)(u,n)v_n=V_t^{(s)}(u,n) is projected into query/key vectors with shared affine mappings, qn=Wqvn+bqq_n=W_qv_n+b_q and km=Wkvm+bkk_m=W_kv_m+b_k. After dropping terms constant within a softmax row, the attention reduces to

At(s)(u,n,m)=softmax⁡m[(avn+b)vm],A_t^{(s)}(u,n,m)=\operatorname{softmax}_m[(a v_n+b)v_m],

where a,ba,b depend on the shared projection parameters and attention scaling. This is an input-dependent but restricted interaction: it contains no explicit variable-identity embedding or freely learned pair-specific dependency parameters. Although upstream features may already contain cross-variable information, they do not remove this representational limitation. We will therefore call the matrices attention-based relation proxies rather than recovered physical or causal dependency graphs.

  1. Normal versus anomalous differences. Temporal pooling, resolution changes, and legitimate operating-state variation can alter these matrices under normal conditions. Conversely, anomalous states may produce similar proxies across scales. Normal-data consistency learning provides an inductive reference, not a guarantee of distinguishing genuine dependency changes from ordinary scale effects. SrelS_{\mathrm{rel}} is therefore auxiliary evidence, not direct proof of a dependency change.

  2. Why penalize differences? The loss encourages, but does not enforce, similarity across scales. Our statement that matrices need not be identical means that no hard equality constraint is imposed; it does not mean the loss leaves differences unaffected. We evaluated αrel=0\alpha_{\mathrm{rel}}=0 versus the default 0.0050.005:

Datasetαrel=0\alpha_{\mathrm{rel}}=0αrel=0.005\alpha_{\mathrm{rel}}=0.005Gain
NIPS-TS-Water0.90200.9102+0.0082
SMD0.93850.9423+0.0038
SWaT0.97910.9816+0.0025
TAO0.93870.9443+0.0056
Mean PA-F10.93960.9446+0.0050

The average improvement is 0.50 percentage points under these settings, but is modest and not established as statistically significant by a paired test. The loss may also suppress meaningful normal or anomalous scale differences. We will make both the limited semantics and this trade-off explicit.

Reviewer_igKkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CHIME, a framework for multivariate time series anomaly detection that addresses the limitations of reconstruction-based methods. It adds historical context to bidirectional cross-scale modeling and introduces two inconsistency measures. These measures are combined with reconstruction errors to produce anomaly scores. Experiments on 10 public benchmarks show strong point-adjusted F1 performance, although the method's novelty compared with existing multi-scale anomaly section methods needs further clarification.

Critical strengths

S1: The paper provides a clear and well-motivated approach to structural anomalies beyond reconstruction errors. It considers an important limitation of reconstruction-based time-series anomaly detection that anomalous observations may look normal in value while breaking time or variable relationships. It directly measures time inconsistency between alignment representation at adjacent resolutions, and inconsistency between variable-relation structures across resolutions.

S2: Extensive experimental evaluations are performed. The paper evaluates CHIME on 10 public benchmarks covering diverse application domains and dimensionalities.

Critical weaknesses

W1: It is not very clear what the novelties of CHIME are compared with existing context-aware and multi-scale methods. The paper claims that existing methods mainly use multi-scale representations to enrich features, while CHIME directly measures differences across resolutions. However, this difference is not developed enough to show a clear conceptual advance over closely related methods such as CrossAD, which already uses cross-scale associations and cross-window context. CHIME combines historical context conditioning, bidirectional cross-scale attention, progressive reconstruction, and two inconsistency-based scores, but it does not clearly separate what fundamentally makes its inconsistency modeling different from existing cross-scale association or reconstruction mechanisms. As a result, the method may look like a combination of established components with extra discrepancy-based objectives instead of a clearly different methodological contribution. The authors are suggested to show clearly the novelty boundary of this work compared with other existing approaches.

W2: The main assumption that cross-scale inconsistency is a reliable sign of anomalies needs stronger support. CHIME treats differences between adjacent-scale latent representations, decoder output, and variable-relation structures as evidence of anomalies, and encourages consistency through auxiliary training objectives on normal data. However, representations at different time resolutions are not necessarily expected to agree. Downsampling and temporal aggregation can legitimately change local patterns and inter-variable dependencies, even under normal operating condition. Therefore, the proposed discrepancy scores may reflect inherent scale-dependent variations or representation artifacts instead of anomalous behavior. On the other hand, real anomalies may appear consistently across resolutions and thereby producing only limited inconsistency scores. The paper does not sufficiently establish when cross-scale differences indicate abnormality rather than normal temporal variation.

Questions for authors

Q1: What is the fundamental distinction between CHIME's inconsistency modeling and existing cross-scale association modeling?

Q2: How is the anomaly threshold τ\tau selected, and how the threshold-selection rules are applied to all baselines?

Q3: What is the reason for directly comparing latent representations at different time resolutions? In Equations (6) and (7), CHIME measures temporal inconsistency by comparing latent representations and decoder outputs from adjacent scales after temporal interpolation. However, representations at different resolutions may naturally encode different temporal meanings, and interpolation only aligns their time dimensions without necessarily making them semantically comparable. What properties of architecture or training objective ensure that these cross-resolution differences reflect meaningful structural inconsistencies rather than natural differences between scale-specific representations? More broadly, why should such differences be seen as evidence of anomalies?

3 comments
AuthorsOpenReview ↗

Response to Q1: Distinction from Existing Methods

Thank you for asking us to clarify the conceptual distinction. We agree that cross-scale association modeling is well established and that CHIME does not introduce cross-scale interactions for the first time. Nor should its contribution be characterized as merely combining multi-scale features, historical context, and reconstruction.

The distinction concerns how cross-scale information is used for anomaly detection. Methods such as CrossAD exploit associations across resolutions to facilitate representation learning and reconstruction. CHIME additionally treats the residual differences between adjacent scales, after context conditioning and bidirectional interaction, as explicit anomaly evidence:

  1. Temporal inconsistency compares time-aligned latent representations and decoder outputs across adjacent scales.
  2. Variable-relation inconsistency compares scale-specific, attention-based relation proxies.
  3. Both components are constrained by auxiliary consistency objectives on normal training data and contribute directly to the inference score alongside reconstruction error.

Thus, cross-scale association and inconsistency are complementary rather than mutually exclusive: interaction exchanges information across resolutions, whereas inconsistency scoring measures what remains discrepant relative to the normal cross-scale responses learned during training.

We also acknowledge that association discrepancy has been used as an anomaly signal, notably by Anomaly Transformer, and that CrossAD already models cross-scale associations. We do not claim novelty for structural discrepancy in general. The more specific contribution is the joint use of context-conditioned, adjacent-scale temporal and relation-proxy discrepancies as explicit auxiliary scores in a hierarchical reconstruction framework. We will sharpen this distinction and avoid overstating the novelty boundary in the revision.

AuthorsOpenReview ↗

Response to Q2: Threshold Selection and Fair Evaluation

Thank you for identifying the missing threshold details. We audited the evaluation code and distinguish the protocols used for different metrics.

For the reported point-adjusted F1 (PA-F1), precision, and recall, the procedure is retrospective and uses test labels. Ground-truth anomaly intervals are used for point adjustment: scores within an anomalous interval are replaced by the interval maximum. We then apply precision_recall_curve to the adjusted test scores and labels and select the threshold maximizing F1. Each method has its own score distribution and hence its own optimal threshold, but all ten baselines were re-evaluated using the same point-adjustment, threshold-generation, and F1-maximization procedure. This is an oracle-style test-label-assisted evaluation, not a label-free deployment protocol.

For the reported Affiliation F1 (AF-F1), the evaluation setting named optimal uses the 99th percentile of the unadjusted test anomaly scores as the binary threshold, followed by event construction and Affiliation evaluation. The percentile itself is not selected using test labels, but it does use the test-score distribution; it is therefore not a threshold calibrated on an independent normal validation set. The code also supports a separate training/validation-based preset option, which should not be confused with the reported optimal setting.

AUC-ROC, VUS-ROC, and VUS-PR are computed from continuous scores and do not require a single operating threshold. We will add an explicit protocol description covering the threshold source, use of test labels, point adjustment, and common baseline evaluation procedure for each metric.

AuthorsOpenReview ↗

Response to Q3: Cross-Scale Representation Comparability

We agree that temporal interpolation aligns time indices but does not by itself make latent representations at different resolutions semantically equivalent. CHIME does not assume unconditional equality between scale-specific features. Rather, it learns a correspondence under normal operation and uses deviations from that learned correspondence as an auxiliary anomaly cue.

All scales are derived from the same input window by different degrees of temporal pooling and retain the same variable identities. A shared scale-encoding mechanism and bidirectional cross-scale attention enable coarse trends and fine-grained responses to condition one another. Reconstruction and auxiliary temporal-consistency losses on normal training data further encourage relatively stable relationships between adjacent-scale latent features and decoded outputs. Consequently, Eqs. (6)–(8) measure residual discrepancies after learned interaction and temporal alignment, not arbitrary distances between two independently trained representations.

This provides an empirical rationale, not a formal guarantee of semantic equivalence. Normal high-frequency details, pooling, and aggregation can legitimately create cross-scale differences; conversely, some anomalies may affect all scales similarly and yield small discrepancies. Comparing decoder outputs in a common observed-variable space is more directly interpretable than comparing latent features, but it is not free from this limitation.

We therefore do not regard cross-scale discrepancy as a necessary or sufficient condition for an anomaly. It is one source of evidence, combined with multi-scale reconstruction error and the relation-proxy discrepancy. We will revise the manuscript to describe these quantities as learned, normal-reference inconsistency proxies rather than theoretically guaranteed indicators of structural anomalies.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.