acceptodds
Under review as a conference paper at ICLR 2027

SCAST: Spectral Constraint-Aware Selective Transfer for Cross-System Log Anomaly Detection

Abstract

Cross-system log anomaly detection transfers knowledge across systems with different event vocabularies and interaction structures, often under limited target-domain supervision. A key challenge is deciding which structural dependencies should be transferred before reducing source–target discrepancy. We present SCAST, a spectral constraint-aware selective transfer framework that performs dependency selection before alignment. SCAST maps heterogeneous log graphs into a shared semantic–structural space and decomposes their structural signals into multiscale spectral bands. Within each band, temporal direction and structural priors guide the modeling of dependency candidates, while a learned router separates features used for transfer from residual information. The selected band-wise representations are then fused into sample-level summaries for cross-domain alignment and anomaly scoring, rather than aligning an entangled whole-graph representation. We analyze the effects of cross-band fusion and the trade-off between reducing domain discrepancy and retaining predictive information.Experiments on four public log datasets show that SCAST performs competitively in cross-system transfer while remaining effective in single-system detection. Component and alignment-placement comparisons indicate that screening dependencies before alignment is more effective than aligning mixed representations. In an extended diagnostic model, frequency-resolved perturbations produce larger mean anomaly-logit changes for the screened branch than for residual and random perturbations in each of the three tested structural bands.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Log Anomaly Detection, Pattern Mining

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_9vBzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCAST, a framework for cross-system log anomaly detection under heterogeneous event vocabularies, graph structures, and limited target-domain anomaly supervision. The main idea is to replace direct alignment of mixed source and target representations with a decompose–screen–transfer pipeline. SCAST first constructs graph-based representations of log windows and decomposes their structural signals into multiple graph-spectral bands. Within each band, spatial and temporal priors are used to model candidate dependencies. A feature-wise router then separates the representation into a screened branch and a residual branch. Only the screened branch is dynamically fused across spectral bands, used for anomaly scoring, and subjected to cross-domain adversarial alignment. The paper motivates this design by arguing that domain shifts may differ across structural frequency bands and can partially cancel after premature feature fusion. The authors also provide a conditional domain-adaptation bound characterizing when selective alignment can improve a target-risk upper bound. Experiments are reported on HDFS, BGL, Thunderbird, and OpenStack. The paper includes single-system results, historical cross-system comparisons, matched HDFS↔OpenStack controls, component ablations, frequency-resolved perturbation analyses, efficiency studies, and spectral-basis sensitivity analyses.

Critical strengths
  1. The paper identifies a meaningful limitation of indiscriminate domain alignment. The distinction between aligning a mixed representation and first screening potentially transferable information is conceptually useful. The matched alignment-placement comparison provides direct evidence that the ordering of screening and alignment matters in the evaluated setting.
  2. The empirical study is reasonably thorough and includes useful diagnostics beyond headline performance. The paper includes component ablations, matched target-only and pre-alignment controls, spectral-band analyses, perturbation-based sensitivity diagnostics, efficiency measurements, and spectral-basis sensitivity tests. I particularly appreciate that the paper explicitly reports limitations of several analyses rather than overclaiming them.
Critical weaknesses
  1. The central notion of a “transferable dependency” is not independently identified or validated. The paper’s core scientific interpretation is that SCAST screens out system-specific dependencies and retains dependencies that are more suitable for cross-system transfer. However, there is no independent supervision or ground truth indicating which dependencies are actually transferable. Operationally, the router is trained using the same anomaly-prediction, structural regularization, and domain-alignment objectives that define the final task. The screened branch is then the only branch used for anomaly scoring and alignment. Therefore, what is directly established is that the optimization procedure learns a representation that is useful for the evaluated prediction and alignment objectives. This does not by itself establish that the screened branch corresponds to genuinely stable cross-system structural dependencies. The perturbation analysis does not fully resolve this issue. It shows that perturbing the screened branch causes larger anomaly-logit changes than perturbing the residual or random branches in the tested configuration. This establishes greater predictive sensitivity, but predictive importance is different from transferability. A feature can strongly affect anomaly prediction while still being system-specific. This distinction is important because “selective transfer of transferable dependencies” is the main conceptual claim of the paper. Without an independent notion of transferability, the interpretation is substantially stronger than the evidence currently supports.
  2. The learned spectral dependency representation is basis-dependent, which weakens its interpretation as a stable structural object. The appendix shows that the tested predictors are not invariant to eigenvector sign flips or rotations within repeated eigenspaces. These transformations can correspond to mathematically equivalent representations of the same eigenspaces, yet they produce non-negligible and sometimes large changes in model logits. This is particularly concerning here because the method interprets learned quantities in spectral coordinates as frequency-resolved structural dependencies. If model behavior depends materially on an arbitrary choice of spectral basis, then it is unclear whether the router is discovering stable graph-structural relationships or basis-specific coordinate patterns. The authors correctly note that the reported logit changes do not directly measure F1 degradation, but I do not think this eliminates the conceptual issue. The lack of basis invariance directly affects the interpretation of the learned spectral representations, which is central to the paper’s claimed mechanism.
Questions for authors
  1. Can the authors directly validate transferability rather than predictive importance? Is there an experiment in which dependencies selected in one source–target pair are evaluated for stability or usefulness in a different target system, without retraining the router for that pair? For example, if the screened representation truly captures transferable dependencies, I would expect it to generalize more consistently across multiple independently chosen target domains than the residual branch. Such an experiment would materially strengthen the central claim.
  2. How should the learned spectral dependencies be interpreted given the demonstrated basis sensitivity? Can the authors construct a basis-invariant or basis-canonicalized version of the method and show that the main transfer results remain intact? Alternatively, can the authors quantify whether the observed sign-flip and eigenspace-rotation sensitivity affects actual ranking metrics such as PR-AUC or ROC-AUC, rather than only logits? A convincing answer here would materially affect my assessment.
  3. How much of the gain comes from selective alignment itself, versus the particular spectral decomposition? The current results support the full pipeline, but the broader scientific idea seems to be “screen before align.” Can the authors test whether the same principle holds with a non-spectral decomposition or another representation family? If the advantage persists, that would support a more general selective-alignment principle rather than a method tied specifically to this graph-spectral parameterization.
  4. Can the authors clarify the strength of the cross-system evidence under matched protocols? The historical cross-system table mixes results from methods with heterogeneous preprocessing, data splits, label budgets, encoders, and thresholding protocols. The most controlled comparisons appear to be concentrated in HDFS↔OpenStack. How much of the claimed cross-system generality remains if the conclusions are restricted to only matched-protocol experiments?
3 comments

Responses to Question 1: In order to validate transferability, we have specially designed the following experiment: to froze the representation, screening, routing, and selection parameters after training on HDFS → OS, then evaluated selected, residual, and gate‑location‑shuffled representations on BGL and TB. Identical linear readouts use only OS labels and share feature dimension, optimizer, stopping rule, and evaluation windows. External‑target access is limited to predefined normal reference data for graph construction and evaluation, with no parameter updates. Results are summarized below:

Unseen targetSelected PR‑AUCResidual PR‑AUCShuffled PR‑AUCSelected − ResidualSelected − Shuffled
BGL0.812 ± 0.0120.711 ± 0.0090.740 ± 0.016+0.101 [0.082, 0.125]+0.072 [0.051, 0.093]
TB0.786 ± 0.0100.668 ± 0.0170.706 ± 0.015+0.118 [0.089, 0.146]+0.080 [0.058, 0.104]

The selected representation achieves higher PR‑AUC than both controls on both unseen targets. All descriptive paired 95% intervals are above zero, supporting the predictive utility of the learned selection locations beyond the training pair. Furthermore, we have added a new Section 4.6 dedicated to the discussion of transferability.

Responses to Question 2: In order to verify the basis sensitivity, we extended the analysis to complete target test sets, using three independently trained checkpoints per direction and 20 pre‑specified transformations per type (Table 21, Appendix G.2). Graphs, inputs, learned parameters, and validation-selected thresholds were fixed. Results indicate that the proposed model largely preserves its average ranking performance under the tested spectral-basis transformations, providing empirical support for its task-level stability:

DirectionTransformationPR‑AUC, mean ± SDROC‑AUC, mean ± SDF1, mean ± SDMean ΔPR‑AUCLargest paired PR‑AUC drop
HDFS → OSSign flip0.8972 ± 0.01100.9440 ± 0.002300.9560 ± 0.00230−0.00320.0200
HDFS → OSEigenspace rotation0.9012 ± 0.00950.9490 ± 0.01400.9670 ± 0.0210+0.00080.0100
OS → HDFSSign flip0.999200 ± 0.0004000.999950 ± 0.0000300.9945 ± 0.0045−0.0007650.0045
OS → HDFSEigenspace rotation0.999800 ± 0.0001000.999988 ± 0.0000080.9950 ± 0.0040−0.0001650.0012

Responses to Question 3: First, to test screening before alignment beyond spectral decomposition, we replaced spectral features with one‑ to three‑hop propagation features in node space and compared screened versus direct alignment across five directions (Table 4, Section 4.5).

DirectionSCAST screened / direct F1Multi‑hop non‑spectral screened / direct F1
BGL → TB0.9600 / 0.89650.933 / 0.6933
HDFS → OS0.9696 / 0.90870.9317 / 0.8099
TB → BGL0.9310 / 0.78960.9286 / 0.4761
BGL → HDFS0.9730 / 0.95450.9564 / 0.9289
HDFS → BGL0.8930 / 0.86790.8884 / 0.4381

Screening improves F1 in both representation families across all five directions.

Second, to test the same principle in an external method, we added screening before alignment to LogTAD (Table 5, Section 4.5). The table reports BGL → TB.

LogTAD configurationF1
Original LogTAD0.7774
LogTAD + screening before alignment0.8823

Adding screening raises F1 by 10.49 percentage points, supporting its benefit in an external alignment model.

Third, to isolate alignment placement while retaining screening, we compared three alignment settings in the multi‑hop non‑spectral model (Table 17, Appendix E.3).

Alignment settingF1
No alignment0.9082
Pre‑screen alignment0.9255
Post‑screen alignment0.9333

Post‑screen alignment achieves the highest F1, exceeding pre‑screen alignment by 0.78 percentage points and no alignment by 2.51 percentage points. This supports aligning the screened representation for thresholded detection.

Responses to Question 4: Owing to inherent differences in model architectures, some protocol components, such as label budgets and encoders are difficult to align perfectly. We adopt LogTAD (with released code) and add a new baseline ARMET, and provide new experimental results (F1$\uparrow$) under adapted protocol (Table 2). It can be observed that SCAST achieves competitive performance with the matched protocol.

MethodBGL → TBTB → BGLHDFS → BGLBGL → HDFSHDFS → OSOS → HDFS
LogTAD0.7770.7100.6580.7070.7210.869
ARMET0.8080.8600.8010.8990.9190.878
SCAST0.9600.9310.8930.9730.9690.996
Reviewer_9vBzOpenReview ↗

Follow-up on the Rebuttal

Thank you for the additional experiments. The frozen-model results on new target systems, along with the non-spectral and LogTAD comparisons, address my main questions about transferability and the broader usefulness of screening before alignment.

Please highlight these results in the revision and be more precise about the claim of discovering transferable structural dependencies. The new stability tests are also helpful, though stable predictive performance should be distinguished from strict spectral-basis invariance.

Thank you for the comments.

We have highlighted the above results in the revised manuscirpt. Meahwhile, we supplemented and revised the relevant content to ensure that the claims are stated more rigorously, comprehensively, and precisely.

As to transferability, we have complemented an entirely Section 4.6 dedicated to the discussion of transferability. And, the newly added transferability experiments and their detailed discussions are included therein.

As to basis sensitivity, we have complemented the experimental results on performance under spectral-basis transformations as well as relevant discussions in Table 21 of Appendix G.2. Furthermore, we explicitly state that the experiments provide empirical support for its task-level stability of predictive performance, rather than strict spectral-basis invariance.(In the the preceding paragraph of Appendix G.2 of the revised manuscript)

As to selective alignment, we have complemented an entirely Section 4.5 dedicated to the discussion of “screening before alignment”. The newly added experiments and their detailed discussions are included therein.

As to comparisons under matched protocols, we have complemented the related experimental results in Table 2.

All the above revisions are highlighted in blue in the revised manuscript.

Reviewer_BxAjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCAST, a spectral constraint-aware selective transfer framework for cross-system log anomaly detection. The method first decomposes log graph representations into multiple spectral bands, then uses temporal and structural priors to screen potentially transferable dependencies while routing complementary information into a residual branch. The selected representations are dynamically fused and aligned across source and target systems before being used for anomaly scoring. Experiments on four public log datasets evaluate SCAST in both single-system and cross-system settings, with ablation studies and additional analyses examining the effects of selective alignment, structural regularization, and multiscale representation learning.

Critical strengths
  1. The paper tackles an important problem in cross-system log anomaly detection. The idea of screening potentially transferable dependencies before domain alignment is well motivated, especially when source and target systems differ substantially in their graph structures.

  2. The proposed decompose–screen–transfer framework is intuitive and well structured. Combining multiscale spectral representations with constraint-aware screening and selective alignment offers a reasonable way to address the challenges of cross-system transfer.

Critical weaknesses
  1. The framework combines several existing techniques, including graph spectral filtering, spatiotemporal attention, feature-wise gating, adversarial domain alignment, and standard regularization terms. While screening dependencies before alignment is a reasonable design choice, it remains unclear what makes this combination substantially different from existing approaches and whether each component is necessary. The paper would benefit from a clearer explanation of its methodological contributions and stronger comparisons showing that the proposed design offers advantages over simpler alternatives.

  2. The overall objective in Eq. (20) combines several standard loss terms, including anomaly supervision, adversarial domain alignment, structural regularization, and orthogonality constraints. However, the paper provides limited justification for how these objectives interact or how sensitive the results are to their relative weights. A more detailed analysis is needed to demonstrate that the combined objective is well motivated rather than simply an aggregation of existing losses.

Questions for authors

See Weaknesses.

1 comment

Responses to Weakness 1: According to this comment, we have complemented the following experiments to verify the contribution and necessity of the components (Due to character limits, more detail experiment setups and discussions have been provided in the revised manuscript).

First, to assess the main design choices, we have conducted additional component ablations on BGL → TB (Table 3,Section 4.3).

Component examinedAblation / replacementF1 ↑\uparrow
Full modelFull SCAST0.960
Multi‑band representationWithout multi‑band0.927
Constrained attentionGeneric attention0.926
Structural regularizationWithout structural regularization0.854
Dynamic fusionWithout dynamic fusion0.922
Domain alignmentWithout domain alignment0.855
Alignment placementPre‑screen alignment0.925
Source‑transfer pathwayTarget‑only0.906

Second, to compare the full design with a simpler alternative, we added the experiments by reporting a BGL → TB variant using generic attention and uniform band fusion (Table 15, Appendix E.3).

ConfigurationF1
Full SCAST0.9451
Generic attention + uniform fusion0.9242

Third, to test screening before alignment across model designs, we compared SCAST, a multi‑hop non‑spectral model, and LogTAD (Tables 4 and 5, Section 4.5).

ModelWithout screening before alignment: F1With screening before alignment: F1
SCAST0.89650.9600
Multi‑hop non‑spectral0.69330.9330
LogTAD0.77740.8823

Screening improves F1 for all three models on BGL → TB and for both SCAST and the multi‑hop model across all five tested directions. LogTAD AP also increases from 0.7919 to 0.8326. This supports screening before alignment beyond the spectral representation; And, Tables 15 and 17 provides complementary controls.

Responses to Weakness 2: In order to verify the effectiveness of combined objective, we have supplemented the following experiments for validation.

First, to test whether each loss contributes to detection, we removed one term at a time on BGL → TB. Table 3 in Section 4.3 reports the results summarized below:

BGL → TB variantF1Change from Full (percentage points)
Full SCAST0.960—
Without sparsity0.922−3.8
Without orthogonality0.914−4.6
Without structural term0.913−4.7
Without alignment0.895−6.5

Each removal lowers F1, with the largest decrease after removing alignment. Supplementary ablations in Table 12, Appendix D.1 report AP on BGL → TB.

Second, to test alignment–orthogonality interaction, we switched two terms on and off while retaining source training and both branches (Table 13, Appendix D.2). The table reports mean AP on HDFS → OS.

AlignmentOrthogonality offOrthogonality onAP change
On0.90210.9298+0.0277
Off0.81260.8032−0.0094

Orthogonality improves AP with alignment but lowers it without alignment, indicating a descriptive interaction between the two terms.

Third, to test weight sensitivity, we varied each weight tenfold downward and upward while fixing the other in a separate HDFS → OS reference configuration (Table 19, Appendix F.1).

Alignment weightOrthogonality weightF1AP
0.005 (reference)0.02 (reference)0.92010.9462
0.00050.020.92830.8259
0.050.020.91090.8093
0.0050.0020.86200.8154
0.0050.20.91930.8054

The reference weights give the highest tested AP, whereas a smaller alignment weight gives the highest F1. Table 18 in Appendix F provides complementary BGL → TB results.

Reviewer_oCyXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SCAST addresses cross-system log anomaly detection by screening dependencies within spectral bands and aligning the selected summaries across systems. Anomaly scores are computed from the resulting dependency sequences.

Critical strengths
  • Screening dependencies before alignment is a useful direction for handling structural differences across systems.
  • The controlled comparisons help assess source information and alignment placement, although some gains are small relative to the reported run-to-run variation.
Critical weaknesses
  • Without target anomaly labels, the normal-sample loss allows a constant normal prediction. It is unclear what prevents this. The fixed-protocol experiments use labeled target anomalies and validation data, so they do not establish performance without target anomaly labels.
  • Alignment uses pooled summaries, while detection uses ordered dependency sequences. Aligning the summaries does not necessarily align the temporal information used for detection. The perturbation analysis shows predictive sensitivity, but does not establish transferability.
  • The related work section could include more references on cross-domain graph anomaly detection and domain adaptation. Some relevant papers are Cross-Domain Graph Anomaly Detection via Graph Transfer and Graph Decouple (2025), Cross-Domain Graph Anomaly Detection via Test-Time Training with Homophily-Guided Self-Supervision (TMLR 2025), Cross-Domain Graph Level Anomaly Detection (IEEE TKDE 2024), and Information Filtering and Interpolating for Semi-Supervised Graph Domain Adaptation (Pattern Recognition 2024). It would be useful to discuss how SCAST differs from these approaches.
Questions for authors
  • Is alignment restricted to normal windows, or are labeled target anomalies also included? This matters because source training uses only normal windows.
2 comments

Responses to Weakness 1: Thank you for the reviewer’s comments. The reported experiments use labeled target anomalies during training and a separate labeled target validation set for checkpoint and threshold selection. In the evaluated setting, anomaly supervision penalizes low scores on labeled anomalies, while the ranking term encourages their scores to exceed those of target normal samples. Together, anomaly supervision and ranking penalize the constant‑normal solution. In particular, equal normal and anomaly scores incur the positive ranking margin, while predicting every sample as normal incurs a high anomaly classification loss.

In order to avoid potential misunderstandings, We have revised Section 3.1 to define the labeled target anomaly training set explicitly and distinguish training supervision from labeled validation data. And, we have complemented the direction‑specific training and validation budgets in Table 10, Appendix C.3.

Transfer directionTarget training anomaliesTarget validation anomalies
BGL → TB15190
TB → BGL6,2643,990
HDFS → BGL6,2643,990
BGL → HDFS1111,337
HDFS → OS3233
OS → HDFS32134

Responses to Weakness 2: First, pooled summaries provide compact, fixed‑dimensional alignment interface across systems with different event vocabularies and graph sizes. The detection pathway retains the ordered screened representations and processes them with a temporal convolutional network, as described in Section 3.4. In order to verify whether an order‑aware alignment summary improves detection without replacing the temporal detector, we have added a control test by comparing this interface with Ordered‑align (Table 15, Appendix E). Results are as follows:

DirectionAlignment interfaceF1AP
HDFS → OSPooled screened summary0.96960.9298
HDFS → OSOrdered‑align0.89580.8195
BGL → TBPooled screened summary0.94510.9354
BGL → TBOrdered‑align0.90270.9473

The pooled interface yields higher F1 in both directions. Its AP advantage is direction‑dependent: Ordered‑align has higher AP on BGL → TB. These results support the pooled interface for thresholded detection in both directions.

Second, We add new Section 4.6 that provides a detailed discussion of transferability, where a frozen cross‑target experiment has been implemented. SCAST is trained on HDFS → OS, after which its representation, screening, and routing parameters are frozen. Identical linear readouts use only OS labels, and evaluation is performed on BGL and TB without updating parameters on either external target. This experiment supplies cross‑target evidence beyond the predictive‑sensitivity analysis.

Unseen targetSelected PR‑AUCResidual PR‑AUCShuffled PR‑AUCSelected − ResidualSelected − Shuffled
BGL0.812 ± 0.0120.711 ± 0.0090.740 ± 0.016+0.101 [0.082, 0.125]+0.072 [0.051, 0.093]
TB0.786 ± 0.0100.668 ± 0.0170.706 ± 0.015+0.118 [0.089, 0.146]+0.080 [0.058, 0.104]

Responses to Weakness 3: According this comment, first, we have improved related work section by including the mentioned works. Second, we have added a new Appendix B for more detailed disscusion related to the mentioned works. Third, we have further added ARMET as a baseline across six transfer directions shown in Table 2 (F1$\uparrow$).

MethodB→TT→BH→BB→HH→OSOS→H
LogTAD0.7770.7100.6580.7070.7210.869
ARMET0.8080.8600.8010.8990.9190.878
Full SCAST0.9600.9310.8930.9730.9690.996

Responses to Question: Alignment uses source normal windows and target normal windows. Labeled target anomalies are excluded from the domain‑alignment loss. This matches the normal‑only source training data and avoids using the same adversarial objective to pull target anomalies toward source normal behavior.

The revised Section 3.1 explicitly states that domain alignment uses only normal windows. And, we complemented supervision budgets in Table 10, Appendix C.3.

Reviewer_oCyXOpenReview ↗

Official response from reviewer

Thank you for your response. The rebuttal addresses some of my technical concerns. However, I maintain my original assessment, as the benefits of cross-system transfer, the necessity of the proposed architecture, and the overall methodological contribution remain insufficiently convincing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.