acceptodds
Under review as a conference paper at ICLR 2027

SGALAD: Subgraph-Guided Adaptation for Log Anomaly Detection Across Domains

Abstract

Cross-system log anomaly detection must handle changing event vocabularies, dependencies, and anomaly distributions with few target labels. We present SGALAD (Subgraph-Guided Adaptation for Log Anomaly Detection), which learns local event subgraphs and adapts a frozen graph representation. It selects seed events, refines neighborhoods, and fuses local features through attention. Target adaptation updates a projection, classifier, and domain discriminator while preserving the source-trained extractor. On BGL-to-Thunderbird and Thunderbird-to-BGL transfer, SGALAD achieves 10-shot F1 of and , exceeding size-matched random-subgraph adaptation by 7.90 and 5.50 points. Across 5, 10, and 20 labeled windows per class, it outperforms full GNN fine-tuning. SGALAD updates 0.06 million parameters (4.8% of the full-GNN trainable count), with approximately 8.6–8.8 times shorter measured adaptation time. Fixed-representation discrepancy and fixed-feature estimation bounds motivate both design choices. These results support learned local selection for efficient cross-system adaptation.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Log anomaly detection, Cross-system domain adaptation, Graph neural networks, Local subgraph learning, Parameter-efficient adaptation

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 4 reviews, above 27% of the others.
Reviewer_juZJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a subgraph-guided adaption method for log anomaly detection across domains and the extensive experiments are conducted to show the effectiveness and efficiency across various settings, including single-domain log detection, cross-domain log detection, few-shot setting, and adaptation efficiency.

Critical strengths
  1. This paper proposes a subgraph-guided adaption method for log anomaly detection across domains.

  2. The experimental resutls demonstrate the efficiency during the domain adaptation.

  3. The proposed method outperform most of the baseline methods across multiple settings.

Critical weaknesses
  1. The log constructed graph is usually small. In the paper, the experiments are conducted on Thunderbird and BGL datasets, and according to [1], the number of nodes for these two graphs are 16 and 10, respectively. Many existing works have explored the effectiveness of the graph-based methods. This drives me to question about the necessity of learning which local event structure to retrain for such small graphs. In the experiment, comparison with graph-based baseline methods are not sufficient to support this. The authors mentioned quite a few graph-based log anomaly detection methods in the introduction (e.g., LogGD, Logs2Graphs, and GLAD), while these methods are not compared in the experiment. The authors only compare SGALAD with loggraph in single domain setting and GraphLoRA in cross-domain setting, which is not sufficient to support the central question in the introduction.
  • [1] Zhong Li, Jiayang Shi, and Matthijs van Leeuwen. Graph neural networks based log anomaly detection and explanation. Data Mining and Knowledge Discovery, 40:66, 2026. doi: 10.1007/s10618-026-01235-6.
  1. In the introduction, the authors mainly emphasizes the importance of retaining local event structure and ignores the domain adaptation part. However, the title of this paper highlights the log anomaly detection across domains. The missing emphasis of cross-domain diminished the motivation of anomaly detection across domains. Specifically, different systems may have substantially different event vocabularies, dependencies, and anomaly patterns. It remains unclear why local event structures selected using source-domain supervision would remain informative in an unseen target domain. Since cross-domain adaptation is a central objective of the proposed method, I would suggest the authors to provide a stronger justification for the assumption that anomaly-related local structures are more stable across domains. In addition, the authors do not provide insightful analysis for the experimental resutls in Table 3. Simply mentioning 'Table 3 further examines target performance across explicit support budgets with an expanded baseline set.' is not sufficient.

  2. The proposed framework combines several well-established techniques, including graph-based log representations ([1, 2]), local subgraph selection + attention-based aggregation ([3]), adversarial domain alignment, and parameter-efficient adaptation. Although their integration for cross-domain log anomaly detection is potentially useful, the methodological novelty appears somewhat incremental. This impression comes from the lack of validation of the key research question mentioned in the introduction. Why is it an important research question? In the experiment, the cross-domain experiments mainly involve existing transfer methods (LogTransfer) and general-purpose GNN adaptation approaches (AdaptiveGNN). It remains unclear whether the observed improvements arise from the proposed local subgraph learning mechanism or from the effectiveness of the adaptation framework itself.

  • [1] Zhong Li, Jiayang Shi, and Matthijs van Leeuwen. Graph neural networks based log anomaly detection and explanation. Data Mining and Knowledge Discovery, 40:66, 2026. doi: 10.1007/s10618-026-01235-6.
  • [2] Y. Li, Y. Liu, H. Wang, Z. Chen, W. Cheng, Y. Chen, W. Yu, H. Chen, and C. Liu. GLAD: Content- Aware Dynamic Graphs For Log Anomaly Detection. in 2023 IEEE International Conference on Knowledge Graph (ICKG), Dec. 2023, pp. 9–18. doi: 10.1109/ICKG59574.2023.00007., 2023.
  • [3] Lee, Junhyun, Inyeop Lee, and Jaewoo Kang. "Self-attention graph pooling." In International conference on machine learning, pp. 3734-3743. pmlr, 2019.
  1. In the experiment, the authors compare SGALAD with full GNN fine-tuning and claim its efficency and effectiveness for adaptation in few-shot settings. I am concerned that full finetuning updates for GNN with 1.25M parameters under the limited label information might easily lead to overfitting. The improvement over full GGN finetuning cannot be a strong evidence to support the effectiveness for adaptation in such few-shot settings.

  2. Suggestion for Figure 4: You can add frozen and fire symbols to indicate the frozen neural network and finetuning component. This immediately helps the reader understand which part is frozen and which part can be finetuned.

Questions for authors

Q1: In line 57, you mention that A local graph encoder maps each retained neighborhood to node features rvr_v. Do you mean that rvr_v is derived from puvp_{uv} (the output of equation 3) with the second local GNN?

Q2: In line 81, what is asa_s in Zs~=aszs\tilde{Z_s}=a_s z_s? I assume that asa_s refers to the relevance score of the selected seed node ss, rather than an attention weight generated by the subsequent multi-head attention mechanism. The authors should better clarify this to avoid potential confusion.

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

This paper proposes SGALAD, a subgraph-guided framework for cross-system log anomaly detection under limited target-domain supervision. The method first learns task-relevant local event representations through seed selection, neighborhood refinement, graph encoding, and attention-based fusion. During target adaptation, the pretrained structural extractor is frozen, while a lightweight projection and anomaly classifier are optimized together with adversarial domain alignment. The results show improvements over random-subgraph selection, global representations, and several parameter-efficient adaptation baselines. The paper also provides theoretical discussions based on domain adaptation and fixed-feature generalization bounds.

Critical strengths
  1. The paper reports consistent improvements across multiple target-label budgets and transfer directions. In particular, SGALAD achieves 96.30% and 89.80% F1 under the 10-shot setting on BGL-to-Thunderbird and Thunderbird-to-BGL. The comparisons with size-matched random subgraphs and global representations provide useful evidence that learned local structural selection contributes to the observed improvements.

  2. By freezing the structural extractor, SGALAD updates only 0.06M parameters during adaptation. The reported improvements over full GNN fine-tuning, together with reduced adaptation time and memory consumption, are practically relevant for systems with limited labeled target data.

Critical weaknesses
  1. The paper does not clearly explain how training, validation, and test sets are partitioned across time, or how target-domain validation and threshold selection are performed within the stated label budget. This is particularly important because BGL uses six-hour windows with one-hour strides, creating substantial overlap between adjacent examples. Without an explicit non-overlapping or chronological partitioning protocol, it is difficult to rule out information sharing across evaluation subsets.

  2. The main components of SGALAD, including GCN-based encoding, Top-k selection, edge pruning, attention-based aggregation, frozen-feature adaptation, and adversarial domain alignment, are individually well established. The paper does not sufficiently clarify what new methodological principle emerges from their combination beyond the observed empirical improvements.

  3. Proposition 1 follows from hypothesis-class containment and applies to arbitrary fixed representations; it does not establish that the learned subgraph selection reduces domain discrepancy relative to alternative representations. Proposition 2 concerns norm-bounded linear predictors on fixed features and does not directly cover the nonlinear projection and classifier used by SGALAD.

Questions for authors

Q1

How many labeled target samples are used for validation, early stopping, hyperparameter selection, and decision-threshold selection?

Q2

Are the BGL training, validation, support, and test sets separated chronologically? Given the overlapping six-hour windows, what measures prevent the same raw log events from appearing in different evaluation subsets?

Q3

Can the authors provide cross-domain ablations that independently remove learned seed selection and edge refinement while preserving the adaptation protocol? Is there quantitative evidence that the selected subgraphs are more domain-stable than the original graph representations?

Reviewer_1guDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article proposes SGALAD, which is used to solve the problem of large log differences between different systems and the lack of abnormal labels in the target system. For this reason, the author proposes a two-stage framework. In the source domain training stage, first convert the log into an event diagram, use GCN to code nodes, and then obtain candidate sub-maps through node importance scoring, local neighborhood extraction and edge screening. Finally, use the attention mechanism to integrate these local structures to complete the abnormal detection training. In the target domain adaptation stage, the author freezes the trained graph feature extractor, updates only a lightweight projection module and classifier, and reduces the feature differences between different systems through domain confrontation training. The experiment covers multiple log data sets, as well as the cross-system migration tasks of 5, 10 and 20-shot. The results show that this method is better than the comparison method under most test settings.

Critical strengths

The comparative methods and other settings of this article are relatively comprehensive. The author not only compares with the existing log abnormal detection methods, but also sets up control experiments such as random sub-maps, global map features, complete GNN fine-tuning and removal domain alignment.I think these experiments provide relatively direct empirical support for learning local structures to improve cross-system detection, which is also the most important advantage of this article. However, these results still depend on the subsequent verification of the training mechanism and data protocol. The method takes into account the accuracy rate and target domain adaptation efficiency under the current test conditions, rather than simply obtaining higher performance by increasing the computing volume. This design is of practical significance for scenarios where the abnormal detection model needs to be deployed to multiple systems.

Critical weaknesses
  1. The training mechanism of the edge screening module is not clearly explained, which may affect the technical correctness of the core method. The thesis uses MLP as the side to score, and then keeps the important side through Top-k. However, Top-k is a discrete operation, and the paper does not explain how the edge scoring network obtains the gradient from the final classification loss. The author specifically explains how the node selector is trained through continuous weights, but does not give a corresponding explanation to the edge selector. If the edge scoring network cannot actually be updated normally, then the learning side screening claimed by the paper may not be really realized. This directly affects the credibility of the core method.
  2. Lack of direct verification of the core hypothesis. The paper mainly combines the existing GCN, Top-k selection, Attention, frozen feature extractor and domain confrontation training, and does not put forward a learning mechanism that is obviously different from the existing work. The author believes that local subdigrats can retain stable abnormal information across systems, but at present, the final F1 is mainly used to support this view. Higher detection performance does not directly prove that the selected sub-map does have cross-system stability.
  3. BGL uses highly overlapping sliding windows, but the paper does not fully explain how to avoid the original log overlap between different data divisions. In addition, the number of labels in the target domain verification set and whether they are included in the few-shot budget are not clear enough.
  4. The domain adaptation boundary and fixed feature estimation boundary used in the paper mainly come from the existing theory. It has not been proven that the proposed sub-map selection can reduce the cross-domain difference, nor does it directly cover the nonlinear adaptation network actually used.
Questions for authors

In addition to the above weaknesses, I hope the author can answer: How to convert the numerical time series of SWaT and WADI into the event diagram defined in this article? Did you use additional feature coding or different model structures?

Reviewer_bB4YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses cross-system log anomaly detection in two stages. On the source domain, a graph encoder computes features for each event node; a seed scorer selects top-k seed events, an MLP prunes edges to form local event subgraphs, these are locally encoded and fused via attention, and the whole pipeline is trained end-to-end with source labels. On the target domain, the entire subgraph extractor is frozen, and only a projection layer, an anomaly classifier, and a domain discriminator (with gradient reversal) are trained. Theoretical support comes from two results: a domain discrepancy bound under fixed representations, and an estimation bound under fixed features using Rademacher complexity, justifying limited trainable capacity during target adaptation. Experiments cover single-domain detection (HDFS, BGL, Thunderbird, OpenStack), bidirectional BGL↔Thunderbird and BGL→OpenStack transfer, 5/10/20-shot support sets, internal controls against random subgraphs/global features/full GNN fine-tuning/no alignment, efficiency measurements, and supplementary results on SWaT/WADI.

Critical strengths
  1. The internal controls are well designed. The random subgraph control aligns trainable parameters and the adaptation pipeline to the same level as the full model, changing mainly the node selection mechanism; at 10-shot, the full model outperforms it by 7.90/5.50 percentage points on BGL→TB and TB→BGL respectively, showing that learned selection is genuinely more effective than random selection when alignment is present. The full fine-tuning control changes the optimization scope, the no-alignment control isolates the domain discriminator, and the global features control removes local structure while also reducing adaptation capacity from 0.06M to 0.03M. Compared to papers that only stack external baselines, this factor-by-factor decomposition does a better job of showing where the gains come from.
  2. Parameter count and adaptation cost are real pain points for this task. With 0.06M trainable parameters (4.8% of the full GNN), target adaptation takes 3.83.9 seconds with peak memory of 480485 MiB; even compared to the random subgraph control at the same 0.06M, it adds only about 0.3 seconds and 10 MiB. It outperforms full GNN fine-tuning at all three shot budgets (5/10/20), by 8.80 and 7.20 points at 10-shot in the two directions. If the transfer scenarios can be broadened, this efficiency has practical value for rapid deployment on new systems.
Critical weaknesses
  1. The dependence between the selection gain and domain alignment has not been teased apart. At 10-shot, three cells exist: learned selection with alignment 96.30/89.80, without alignment 88.10/83.40, and random selection with alignment 88.40/84.30; the missing cell is random selection without alignment. The paper reports two conditional effects, 7.90/5.50 (over random subgraph) and 8.20/6.40 (removing alignment), and the numbers themselves are fine. Both, however, are conditioned on the other factor being set to its favorable value. Without the fourth cell, it is impossible to determine whether the two factors interact, or whether learned selection retains any value without alignment. The random subgraph control has nearly the same adaptation cost as SGALAD (3.5s/470MiB vs. 3.8s/480MiB, both at 0.06M), which makes it a clean baseline; but precisely because it keeps full alignment, it cannot answer this question. This cell appears in neither the main text nor the appendix. 2.The coverage and protocol transparency of the cross-domain conclusions are not sufficient. The controlled 5/10/20-shot experiments are limited to BGL↔Thunderbird, both supercomputer logs; BGL→OpenStack is 10-shot one-way only, and more semantically distant transfers such as HDFS or Hadoop have not been attempted. Table 2a juxtaposes SGALAD's 10-shot results with numbers quoted from the original papers for LogTransfer, LogGT, LogETA, and LogDC; Appendix C explicitly acknowledges that these numbers follow each paper's own windowing, splitting, and label-budget protocol, so the 1.6/0.5 point leads cannot be treated as a same-protocol comparison. Table 3 is a more convincing controlled comparison. SWaT/WADI are industrial multivariate time series rather than logs, and the paper does not explain how continuous sensor data are converted into the template event graphs defined here, nor does it provide the splitting and preprocessing details; the results are currently hard to reproduce or interpret. More importantly, the method selection for the main table reinforces an easy-to-misread "highest" claim. Section 5.2 writes that SGALAD "is the only listed method above 96% on all three datasets" and achieves the "highest listed F1" on TB; within the scope of Table 1, both statements hold. However, Appendix Table E.1 shows LogMeta at 99.50/99.30/99.20, LogSnippet at 99.30/98.50/99.60, and LogPDGMoE at 97.11/99.17/99.39, all above 96% on all three datasets, with LogSnippet's TB F1 (99.60) also exceeding SGALAD's 96.78. None of these three methods appear in Table 1; they only appear in Appendix Table E.1. The Table 1 footnote only states that values are quoted from the original papers or the benchmarks of Zhao et al. (2026), Zhang et al. (2026), and Shi et al. (2027), without explaining why these stronger single-domain results were placed in the appendix rather than included in the main comparison. Readers encountering the "highest" claim will not know that higher numbers exist in the appendix; the paper should either state the selection criteria for the main table or qualify the claim to the methods listed therein.
Questions for authors

1.Tables 3 and 4 already show that learned selection outperforms random selection within the full pipeline with alignment. To truly separate the contributions of selection and alignment, please add the random subgraph without alignment condition. To support the "shared anomaly structure" claim beyond just an effective locality prior, please provide concrete examples of top-k seed templates and their neighborhoods in the source and target domains, and quantify how stable the selection results are across runs and support sets. Both points would materially affect the assessment of the mechanistic claim. 2.In Table 2a, SGALAD's numbers are the same as the 10-shot results in Table 3, while Appendix C notes that other methods' numbers follow their respective original protocols. Please list, for each method in Table 2a, the target label budget, unlabeled target data volume, windowing, and splitting. If these cannot be aligned, it would be better to position that table explicitly as background context and limit the main conclusions to the controlled comparison in Table 3. 3.The Related Work section lists MetaLog and LogMeta as few-shot cross-system methods. Are their source domain counts, support set definitions, or task protocols compatible with those used here? If so, why were they not included in the unified comparison; if not, please state the differences explicitly rather than leaving readers to assume direct comparability. 4. Why are there no transfer pairs with greater semantic distance, such as HDFS→BGL or Hadoop→Thunderbird? Both BGL and Thunderbird are supercomputer logs; successful transfer between the two does not rule out the possibility that they are simply similar to begin with.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.