- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
CARNet detects anomalies in text-attributed graphs by comparing attribute and structural community assignments and predicting node attributes from context. It uses cross-community attribute mixing to create training negatives. The method performs well on five synthetic benchmarks, with the largest ablation drop coming from removing these negatives.
Critical strengths
1.The results are competitive against recent TAG anomaly detectors, including CAMERA and CoLL. Separate results for semantic and structural anomalies make the comparison more informative. 2.The ablations help identify where the gains come from, particularly the synthetic negatives and community inference.
Critical weaknesses
1.My main concern is whether the benchmark tests the problem described in the introduction. The injected anomalies change a node’s text to a different topic that is uncommon in its neighborhood. This does not show that CARNet detects conflicts between texts with similar wording, such as the drill/disaster example. The paper also does not test whether legitimate cross-topic connections receive high anomaly scores. 2.I could not determine whether the context predictor has access to the target attributes indirectly. Both the GNN representations and global prototypes are computed from the full feature matrix. Removing the target token from the Transformer does not by itself exclude this information. The separate K-means initialization of the two community heads also needs clarification: how are their community indices aligned before computing KL divergence?
Questions for authors
1.Do you have results on semantically conflicting examples with similar wording, including legitimate cross-topic nodes as controls? 2.Is the target feature masked before computing the GNN representations and prototypes? How are the two sets of community assignments aligned? 3.How were hyperparameters selected without anomaly labels, and what sources of randomness are included in the reported standard deviations?
2 comments
Revisions made in response to your comments are highlighted in orange. Some of the content may overlap with other colored annotations.
A1: The revision adds a more constrained perturbation, with text statistics in Appendix E.1 and Table 6. These rewrites aim to change the topic while retaining the source wording. Across the six datasets, rewrite embedding cosine similarity ranges from 0.7667 to 0.9563, and the retained proportion of distinct original words ranges from 85.91% to 95.71%. These quantities measure resemblance to the source text.
The tables below reproduce the Table 2 comparison of CARNet with CMUCL, CoLL, and CAMERA on Cora, Citeseer, Pubmed, and Wikics under this perturbation setting.
AUROC (%)
Average precision (AP, %)
Within this setting, CARNet has the highest reported AUROC and AP on Citeseer (81.17% and 39.30%) and Wikics (72.23% and 20.04%). Pubmed remains challenging: CARNet obtains 56.20% AUROC and 5.44% AP, while CAMERA and CoLL have the highest reported AUROC and AP, respectively.
Revision and evidence status: Section 5.2/Table 2 and Appendix E.1/Table 6 are present in the manuscript.
A2: The target feature is not masked before computing the GNN representations and community prototypes. Although omitted as a standalone Transformer token, it can influence prediction indirectly through graph propagation and community aggregation. We have clarified this in Section 4.2. The two community heads are initialized independently without explicit index matching. A structure-to-attribute KL regularizer encourages agreement during joint training but does not guarantee initial cluster correspondence. Appendices C–D now clarify this procedure. These revisions describe the existing implementation; no masking or alignment modifications were made.
A3: Hyperparameters were selected through grid search using label-independent criteria. AUROC and AP computed from anomaly annotations were not used to choose the hyperparameter configuration. The candidate values reported in Appendix E.1 are learning rates of {0.001, 0.0005}, community counts of {32, 64, 128}, and neighborhood sizes of {4, 8, 16, 32}.
For CARNet, the standard deviations describe variation across repeated training runs on a fixed constructed benchmark. Section 5.1 and Appendix E.1 now clarify that the training seed controls neural-network initialization, K-means initialization, dropout, and the selection of target and donor nodes for self-supervised feature mixing. The graph, rewritten node texts, encoded features, anomaly annotations, and hyperparameters are held fixed across runs.