- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes TopoDiff, which models memory residuals with topology-conditioned diffusion for anomaly detection in continuous-time dynamic graphs.
Critical strengths
-
Extensive experiments and ablation studies support the method’s effectiveness.
-
The proposed approach seems reasonable overall.
Critical weaknesses
-
The paper does not provide sufficient evidence that memory residuals are more consistent across domains than raw embeddings. Showing actual residual distributions across datasets would help support this claim.
-
Several design choices need clearer justification, particularly inverse spectral projection and the use of diffusion rather than simpler residual models. Eqs. (7) and (8) are differently weighted versions of the same noise prediction error, but their complementary value is not established.
-
The cross-domain results lack analysis. Transfer from MOOC to Bitcoin is weaker than transfer from the other source datasets, while CTDG-SSM is competitive on MOOC-to-Bitcoin-OTC. These differences deserve further discussion.
Questions for authors
- Can the authors show the memory residual distributions for normal and anomalous events across datasets, and compare their cross-domain consistency with raw embeddings?
- The paper distinguishes its approach from snapshot-based graph generation, but the event-induced subgraph appears to be a local snapshot. How is temporal ordering used in its spectral processing?
- Why is inverse spectral projection needed, and what does diffusion offer over simpler residual models? Does combining Eqs. (7) and (8) improve performance over using either alone?
- Why is MOOC-to-Bitcoin transfer weaker, and why does CTDG-SSM perform relatively well on MOOC-to-Bitcoin-OTC? Could the authors evaluate the reverse B to A transfers to examine whether generalization depends on the transfer direction?
- The anomaly scores are evaluated at a fixed set of diffusion steps. How are these steps chosen, and how does the selection balance detection performance against inference cost? Is the same selection strategy used across datasets and in zero-shot cross-domain transfer?
3 comments
Thanks for the author's response, which addressed most of my concerns. I will maintain my positive rating.
We thank Reviewer rWnA for the positive rating and are pleased that our response addressed most of the concerns.
Residual distributions, the inverse projection, transfer direction, and the diffusion-step rule
We thank the reviewer for the precise questions.
For W1 and Q1:
Residual against raw-memory consistency, and the requested distributions.
Answer.
The residual is the consistent representation. Its normal energy keeps a shared range across the six domains while the prior and posterior energies do not.
Evidence and revision.
Appendix A.7.1 adds Table 10 and Figs. 8-9 with the requested distributions, the prior and posterior columns being the raw-memory comparison. Prior and posterior energy are below chance on all six while residual energy is above chance on five, and the anomaly-to-normal ratio runs from 1.45 to 4.29 with Amazon inverting to 0.58. The mean pairwise Wasserstein-1 distance of normal per-coordinate energy is 3.34e-4 for residuals against 1.31e-3 for priors and 1.56e-3 for posteriors.
For Q2:
The subgraph as a local snapshot, and how temporal ordering enters the spectral processing.
Answer.
Temporal order enters through the history-visible edge set, the attention weights that form the node embeddings and the continuous-time encoding in Eq. (5), not through the Laplacian spectrum.
Evidence and revision.
Sec. 4.2 now states this, calls the subgraph a causal snapshot, and notes that the eigenvalues of L_t carry no ordering information.
For W2 and Q3:
Why the inverse projection and diffusion over simpler residual models, and whether Eqs. (7) and (8) are complementary.
Answer.
The inverse is a scale-comparability device rather than an accuracy device, diffusion adds conditioning a fixed residual density cannot express, and Eqs. (7) and (8) are two scalings of one error whose sum matches the better term.
Evidence and revision.
Proof C.1 bounds the spectrum of p(L_t) in an interval independent of the subgraph, which lets the inverse rescale every spectral component by a bounded factor on any graph. Table 11 in Appendix A.7.2 retrains the identity inverse, a Gaussian density and a prior-posterior concatenation, and all three trail the reported model on all six. Table 13 in Appendix A.7.4 reads the components out at step 50, where s1 and s2 differ by at most 0.20 AUROC and their equal-weight sum never exceeds the better term. s2 is retained for scale removal under transfer rather than as a second ranking (Sec. 4.4), s3 is the independent component and leads on Wikipedia, Bitcoin-Alpha and Bitcoin-OTC.
For W3 and Q4:
Why Mooc to Bitcoin is weaker, why CTDG-SSM is competitive there, and whether the reverse transfers were run.
Answer.
One cause explains both observations. The four-dimensional Mooc edge features force the edge projection layer to be rebuilt on every transfer into or out of Mooc, while the HiPPO memory of CTDG-SSM needs no such projection. All eighteen directed pairs were evaluated. The rebuild is symmetric, and the direction comes from the target feature space the reinitialized layer faces. The most direct reading is that a transfer into Mooc faces action features carrying far less information than the 172-dimensional space the other five share, which leaves that layer little structure to recover and gives the weakest group of Table 17 at 49.53, 51.36 and 56.24. A transfer out of Mooc faces that shared space, where the residual normalization of Sec. 4.1 absorbs the scale of a fresh layer, and those three span 56.51 to 72.72.
Evidence and revision.
Sec. 5.3, Appendix B.1, A.7.6 and Table 17 carry this analysis. On Mooc to Bitcoin-OTC, CTDG-SSM reaches 67.51 against 69.43 and takes AUPRC 19.37 against 17.53, the only pair in Table 2 where a baseline leads a metric. Clearing the inherited memory moves that pair to 62.77 while leaving Wikipedia to Bitcoin-Alpha at 73.76, which Appendix A.7.6 reads as the memory carrying part of the transfer once the projection is rebuilt.
For Q5:
How the steps are chosen, the detection-against-cost balance, and whether the strategy is shared.
Answer.
One shared step, 50, fixed a priori for every dataset and every transfer at the midpoint of the cosine schedule where the retained-signal coefficient is about 0.5. The per-event cost is flat across steps, which leaves no trade-off.
Evidence and revision.
Sec. 4.4 and Appendix A.7.4 give the rule, Table 15 lists s1 at steps 10, 25, 50, 75 and 90 with every entry inside a 43.70 to 57.58 band, and Table 16 gives the batched cost, which changes by under 0.01 ms per event between steps 10 and 90. Table 15 is the s1 read-out rather than the final metric, and step 50 was fixed before any target measurement, since choosing the best step per dataset would be a target-side choice the protocol excludes. Appendix A.7.6 states that the scoring weights and the step are inherited by all eighteen transfers.
New in the revision: Appendix A.7.1 to A.7.6 with Tables 10 to 17, Figs. 8-9, and clarifications in Sec. 4.1, 4.2, 4.4 and 5.3.