- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+2)
Read review
Paper summary
AnTAGAD addresses generalist anomaly detection on text-attributed graphs using LLM-selected textual anchors and neighborhood-consistency-based topological anchors. It scores unseen graphs through relative top-k anchor similarity and neighborhood dispersion, and diagnoses over-generalization through anomaly acceptance at matched normal coverage. Experiments on eight target graphs report leading AUPRC on seven targets and leading mean AUROC on all eight, alongside ablations, controlled comparisons, and a conditional separation guarantee.
Critical strengths
- Strong reported transfer performance: AUPRC improves over the strongest baseline by 22.14 percentage points on Pubmed and 15.86 on Texas (Table 1), supporting the practical promise of source-only anchor learning.
- The matched-coverage analysis and frozen-encoder controls (Tables 4–6) provide useful evidence beyond qualitative visualization. The paper also clearly limits the scope of its conditional guarantee.
Critical weaknesses
Update after author response: The authors have substantially addressed my main concerns by correcting Eq. (15), clarifying score normalization and its connection to Theorem 2, and providing capacity-matched ablations and reference-budget details. These clarifications and additional experiments strengthen the support for the method and its reported gains. I am therefore raising my rating from weak rejection to weak acceptance.
Original concerns before the author response:
- The stated method and released implementation differ materially. Equation (15) is not the contrastive probability implemented in code; loss weighting and separately normalized score components also differ from Equations (16)/(18). This leaves the table-producing detector and its correspondence to Theorem 2 unclear, affecting reproducibility and ablation interpretation.
- The anchor contribution remains confounded. Removing TeA/ToA roughly halves bank capacity, while reference counts in Tables 4/6 are unspecified. Appendix J mixes original and injected anomalies without separate results; several injected text corruptions align with the LLM selection criteria. These gaps limit attribution of the gains to instance anchoring.
Questions for authors
- Which code revision/configuration produced the tables? Is Equation (15) a typo, and how do the implemented losses and normalized score correspond to Equations (16)/(18) and Theorem 2?
- Are Tables 4/6 matched in candidate information and per-class storage? Please specify reference counts and clarify whether single-branch gains persist at matched capacity.
- What are the proportions and separate results for original versus injected anomaly types? Please provide the confidence prompts and data/score generation resources needed to verify this benchmark.
We are grateful for the reviewers' constructive feedback. Our responses to Q1, Q2 and Q3 are as follows.
Q1: Generalization to Naturally Occurring Anomalies
A1: We thank the reviewer for raising this important concern regarding the generalization of AnTAGAD to naturally occurring anomalies. We agree that evaluating the model on realistic anomalies is essential for assessing its practical applicability. However, publicly available text-attributed graph (TAG) datasets with naturally occurring, ground-truth anomaly annotations remain limited. To provide a more comprehensive evaluation, we further conduct experiments on representative TAG anomaly detection benchmarks introduced by CMUCL. These benchmarks provide a complementary evaluation setting for assessing the generalization ability of AnTAGAD beyond our original experimental datasets. The experimental results are presented below.
Table 1: Performance Comparison (AUROC/AUPRC)
Q2: Contribution of LLM-Derived Confidence
A2: We sincerely thank the reviewer for raising this insightful question. To further investigate the contribution of instance-level anchor design independently of LLM-derived confidence information, we compare directly using LLM-selected anchors (Anchors) with aggregating the same anchors into K-means prototypes (K-means), while keeping the LLM-derived information consistent. As shown in Table 2, Anchors outperforms K-means in AUROC across all eight datasets and achieves higher AUPRC on seven datasets. These results demonstrate that the performance improvements are not solely attributable to LLM-derived information, but also benefit from preserving instance-level anchor representations.
Table 2: Performance Comparison Between K-means Prototypes and Instance-level Anchors (AUROC/AUPRC).
Q3: Contribution of Neighborhood Dispersion
A3: We thank the reviewer for highlighting the need to clarify the contribution of neighborhood dispersion. To address this concern, we conduct additional experiments comparing anchor-only, dispersion-only, normal-bank-only, anomaly-bank-only, and the full scoring model. As shown in Table 3, the full model achieves the best average AUROC, AUPRC, and OG@95 across eight datasets, demonstrating that neighborhood dispersion complements anchor-based scoring to improve anomaly detection and reduce anomaly acceptance.
Table 3: Average Performance of Different Components (AUROC/AUPRC/OG@95)