- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper presents LORE, a label-free framework for graph anomaly detection that frames node anomaly detection as an evidence-guided risk-allocation problem. The method constructs a graph-adaptive prior from six hand-crafted anomaly charts and derives two self-supervised, node-level evidence signals. It applies two optimal transport procedures to refine node rankings locally and within the high-risk tail, respectively. Experiments on the evaluated datasets report improvements over the baselines.
Critical strengths
- Experiments show improvements on the given datasets compared to the given baselines.
- The risk-allocation interpretation is intuitive.
Critical weaknesses
- The OT component appears close to sorting, which weakens the novelty. Both OT problems use one-dimensional scalar evidence, uniformly weighted one-dimensional targets, equal source/target masses, and squared Euclidean cost. In this setting, the claim of "multiscale optimal transport" seems overstated.
- Most of the empirical gain appears to come from the handcrafted prior, not the proposed OT mechanism, as shown in Table 4. The six-chart prior achieves almost all of the AP, while the entire proposed risk-reallocation machinery adds only a little.
- The prior contains several manually chosen structural assumptions and numerical constants without a clear selection protocol.
- The scalability evidence is incomplete. The authors should analyze the computational cost, including memory and time, for datasets such as T-Social.
Questions for authors
- How are the hyperparameters, such as kappa, lamabda are selected on new datasets?
- Please report ablations of the six prior charts and controller.
- How robust is the soft-normal-reference mechanism when the initial prior is badly inverted or anomaly prevalence is very high?
2 comments
Thank you again for your feedback. we would be grateful to know whether our responses address your concerns, and we would be glad to discuss any remaining questions in the time left.
We sincerely thank the reviewer for the careful and expert reading.
W1: OT and sorting
Computationally, each stage admits an exact sorting solution within its support. Sorting is the solver, while OT is the model, which fixes which reranking is admissible: evidence only permutes existing risk positions inside a prior cell (resolution K) or the tail (mass m), the two scales of "multiscale", and never moves risk across supports. Conditional OT alone reaches 0.2978 macro AP on Main-10, Partial OT alone 0.3063, and both 0.3161 (Table 4). Given the same inputs, tuned fusion reaches only 0.3034 and oracle fusion 0.3116 (see our response to Reviewer NnTV).
W2: Contribution of reassignment
Reassignment improves the prior on all 50 Main-10 graph–seed pairs, by 2.3× on Enron and 38% on Disney (Table 4). Where the prior fails, OT lifts macro AP from 0.2971 to 0.5195 (Table 6). Results with SCAN, OddBall, or Radar as the prior under the same OT (AP):
OT improves every prior on every graph, by 39–65% in macro AP, so the gain is not tied to our prior.
W3 and Q2: Prior design and ablation
The charts, constants, and controller were fixed on Main-10 and applied unchanged to all other graphs. Prior AP of the controller, single charts, a uniform mixture, and the label-selected oracle chart:
The controller closely matches the oracle chart, while no fixed chart or uniform mixture works across graphs. This includes the last eight graphs, which never entered prior development.
W4: Computational cost
One complete scoring run on one RTX 4090:
Q1: Hyperparameters on new datasets
The prior is frozen with no per-graph tuning, so only the OT configuration is selected. λ=(λc,λf) is the direction of κlin, selected with κ, K∈[2,4096], and m∈(0,1/2] by seed-1 AP and frozen for seeds 2–5. The gain is insensitive to K and m (Figure 12).
Q3: Robustness of the normal reference
We retrained the evidence views under three references (AP):
Our reference improves every graph, and even an inverted reference keeps the macro AP at the prior level, since evidence cannot move risk across supports. At high prevalence, it still lowers anomaly contamination, from 20.0% to 14.6% on Minesweeper and from 21.8% to 13.4% on Tolokers, raising AP from 0.3014 to 0.3553 and matching the prior, respectively. Like every label-free detector, LORE assumes its signal ranks anomalies above normal nodes, so a globally inverted prior would be inherited. Our prior satisfies this on all 20 graphs (AUROC > 0.5).
We thank the reviewer again and would be glad to clarify anything further.