acceptodds
Under review as a conference paper at ICLR 2027

Graph Nonlinear Schrödinger Networks for Anomaly Detection

Abstract

Supervised graph anomaly detection has to serve two frequency regimes at once because camouflaged fraudsters mimic their neighbors while other anomalies clash with their neighborhoods. Any single spectral response therefore fails in some regime. We introduce G-SNLS, a deterministic graph network whose depth is a bank of Schrödinger-type propagators. Each head learns its own damping, dispersion and time scale so that attenuation and phase velocity separate while several regimes coexist. Intensity-dependent coupling makes the flow nonlinear while paired decoders absorb the class imbalance. We prove an exact energy-decay law per step together with closure of the response family under composition and a full-depth state bound. G-SNLS takes the leading average rank on both ranking metrics across eight public benchmarks. Through ablations, we attribute the gains to the intensity coupling and the channel map. Compared with an unconstrained filter of matched degree, the constraint preserves accuracy while it yields a response that cannot diverge with depth and a loss surface whose most curved direction is flatter per unit change in the response.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph neural network, anomaly detection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3A4jOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes G-SNLS, a supervised node-level graph anomaly detector based on damped Schrödinger-type spectral propagators. Each spectral head learns damping, dispersion, and time-scale parameters, and the representations are coupled. The model also uses deterministic graph-context features and two decoders trained with focal loss and a smooth average-precision objective. Experiments show improvements on given datasets compared to given baselines.

Critical strengths
  1. Experiments show improvements on given datasets compared to given baselines.
  2. Theoretical statements are clear.
Critical weaknesses
  1. It is unclear that the contribution is Schrödinger-specific. At matched polynomial order, the unconstrained response reaches almost the same accuracy. Thus, the Schrödinger constraint itself does not appear to provide the predictive gains.
  2. The theoretical stability is stronger in claim than experiments. Proposition 2 proves closure and contraction for the linear spectral propagator. However, the framework uses a learned complex channel map, nonlinear modulation, real projection, and radial clipping between propagation steps.
  3. The hyperparameter tuning cost is high for G-SNLS.
  4. The compared baselines and datasets are not comprehensive enough. They should include more recent baselines, such as XGBGraph [1], CONSISGAD [2], and SpaceGNN [3], and more real-world datasets, such as T-Social [1] and DGraph-Fin [1].

Reference

[1] Jianheng Tang, Fengrui Hua, Ziqi Gao, Peilin Zhao, Jia Li. GADBench: Revisiting and Benchmarking Supervised Graph Anomaly Detection. NeurIPS 2023.

[2] Nan Chen, Zemin Liu, Bryan Hooi, Bingsheng He, Rizal Fathony, Jun Hu, Jia Chen. Consistency Training with Learnable Data Augmentation for Graph Anomaly Detection with Limited Supervision. ICLR 2024.

[3] Xiangyu Dong, Xingyi Zhang, Lei Chen, Mingxuan Yuan, Sibo Wang. SpaceGNN: Multi-Space Graph Neural Network for Node Anomaly Detection with Extremely Limited Labels. ICLR 2025.

Questions for authors
  1. Can the authors report AUPRC after removing the ranking decoder and train the strongest baselines with the same focal + smooth-AP objective?
  2. How do results change under equal hyperparameter-search budgets?
4 comments

We thank the reviewer for the precise summary and for the questions, which we answer with new experiments.

W1. Is the contribution Schrödinger-specific, given that an unconstrained response of matched order reaches almost the same accuracy?

This is the finding of Section 4.4, and we state it as such: at matched order the free filter changes AUROC by a mean of −0.09%-0.09\% and the cross-dataset test returns p=0.69p=0.69. The paper does not claim that the constraint raises accuracy over a free filter of matched degree; the ablations attribute the accuracy to the dissipative coupling and the channel map, and the contribution we claim for the Schrödinger parameterization is what the same control shows it to supply at no cost in accuracy:

  • Capacity. Four physical parameters per head shared across layers replace KK complex coefficients per head and layer (KK between 10 and 18 here), and the free filter only matches us when it starts from our response: from random coefficients it loses between 0.47% and 4.93% on all six datasets (p=0.03p=0.03 across datasets).
  • Boundedness. Given order 32 the free filter collapses on three of ten inj-Flickr splits and loses 2.54% and 1.64% uniformly on Minesweeper and Tolokers. The constrained response cannot leave the unit ball at any width or depth (Theorem 1, Proposition 2), and imposing that bound on the free filter removes the collapse (Table 17).
  • Conditioning. The constrained response has the smallest peak curvature per unit change in the response of any arm on all six datasets, by factors of 24 to 2.7×1042.7\times10^4 (Appendix I).

The new experiments add two facts to this picture. The unitary propagator of Kiani et al. inside our architecture, which keeps the complex map and drops the damping, loses 1.31% on Minesweeper, 1.24% on Tolokers, 6.13% on Questions, 0.42% on Elliptic and 1.01% on inj-Flickr with the ReLU of Kiani et al. in place of the modulation, and 3.70%, 2.21%, 6.71%, 0.46%, 1.43% and 2.15% (Amazon included) without any modulation, so the damping is a working part of the parameterization and not a label; and the standard-alternative rows of Table 5 show which forms matter: a real linear channel map stays within ±0.83\pm0.83% of the complex one on every dataset, whereas a ReLU in place of the dissipative modulation costs up to 7.63%.

W2. The stability claim is stronger than the experiments, since the framework also has a channel map, a modulation, a real projection and a clip.

Theorem 3 is the result for exactly that composition. Its proof applies the propagator, the channel map, the modulation, the real projection and the radial clip in the order the layer uses them and gives ∥ψT∥F≤min⁡{BNd, [m(1+ε)]T∥ψ0∥F}\|\psi_T\|_F \le \min\{B\sqrt{Nd},\,[m(1+\varepsilon)]^T\|\psi_0\|_F\} with m=∥M∥2m=\|M\|_2, so the one factor that can expand the state is the channel map, and Proposition 6 bounds what the modulation emits for any input. Proposition 2 is a statement about the propagator alone, and we now say so where it is stated (Section 3.3, blue). We also added Appendix L, which trains the validated configuration of every dataset and reads the state norm after every layer: the state stays within a factor 1.7 of the encoded state on six datasets, within 3.4 on inj-Flickr and grows to 42 over eight layers on inj-Amazon, against bounds between 81 and 5×1075\times10^7; the radial clip touches no entry and the largest entry leaving a modulation equals the ceiling of Proposition 6 to three decimals on every dataset (Table 23). The bound is loose, as the reviewer suspects, but the composition it covers is the complete layer, and the trained models stay far inside it without the clip ever acting.

W3 and Q2. Hyperparameter cost and equal search budgets.

G-SNLS screens forty-eight sampled configurations where a baseline screens eighteen plus its published default, because its space contains discrete architectural choices. Table 22 (Appendix K) answers the question directly with two reruns on the ten splits. Under the baseline budget, the same two-stage search with eighteen sampled configurations screened with the full training length, G-SNLS reaches 85.56, 75.25, 98.07, 97.03, 99.19, 95.09, 97.42 and 95.78% AUROC on Tolokers, Questions, Amazon, T-Finance, Elliptic, inj-Amazon, Minesweeper and inj-Flickr, which keeps the average ranks of 1.12 (AUROC) and 1.00 (AUPRC) of the submitted selection and exceeds it on Minesweeper and inj-Amazon, where screening for the full training length found settings that the shorter screening of the larger search had missed; the main tables keep the submitted selection. Under one shared configuration for every dataset and no search at all, it reaches 85.34, 72.90, 97.57, 97.12, 99.03, 91.50, 94.89 and 94.44, within one point of the validated selection on six datasets, 3.9% below on Questions and 1.3% above on inj-Amazon, and ranks 2.12 on AUROC and 1.50 on AUPRC among the nine baselines. The per-epoch cost of the method is in Appendix M: it is the slowest of the five methods sharing its stack on the sparse graphs and the fastest on Amazon, with less memory than GCN, GraphSAGE and AMNet on the two dense graphs.

W4. More recent baselines and larger datasets.

Baselines. We added XGBGraph (and RF-Graph), ConsisGAD and SpaceGNN, together with PMP, MagNet, UniConv and fLode, on the same ten splits (Appendix G, Table 11). ConsisGAD and SpaceGNN run their released code and configurations; both were designed for the regime of very few labels, so we note that the 40% training split is outside their intended setting. ConsisGAD reaches 77.09, 75.18, 97.82, 93.28, 96.90, 89.45, 88.05 and 90.68% AUROC and SpaceGNN 81.00, 77.14, 97.86, 96.49, 98.88, 95.13, 90.27 and 93.72% on Tolokers, Questions, Amazon, T-Finance, Elliptic, inj-Amazon, Minesweeper and inj-Flickr, against 85.11, 76.78, 98.01, 97.01, 99.23, 90.18, 95.55 and 95.08% for G-SNLS. SpaceGNN passes G-SNLS on Questions (+0.36%) and inj-Amazon (+4.95%) and trails elsewhere by 0.15 to 5.3%; ConsisGAD trails everywhere. XGBGraph is the strongest addition: it leads on Amazon, T-Finance and Elliptic by 0.30 to 0.40% and on inj-Amazon by 8.7%, while G-SNLS keeps the best AUROC on Tolokers and Minesweeper, resolved against all eight additions, and the best AUPRC on Tolokers, Questions, Minesweeper and inj-Flickr.

Datasets. We added T-Social, the social graph released with T-Finance (5.78 million nodes, 73 million edges), on its ten stored GADBench trials, with every method at the configuration it validated on T-Finance and no further search (Table 15, Appendix G). G-SNLS reaches 99.68% AUROC, 97.85% AUPRC and 97.07% macro F1, against 99.73 / 97.88 / 96.82 for RF-Graph, 99.40 / 94.62 / 94.79 for XGBGraph, 96.29 / 70.28 / 83.08 for BWGNN and 84.06 / 22.78 / 60.57 for GPR-GNN, 96.74 / 73.88 / 84.49 for GHRN, 96.23 / 71.83 / 83.21 for GCN and 65.96 / 4.61 / 49.82 for GraphSAGE, whose T-Finance configuration collapses to the all-negative predictor on this graph. The picture of the eight benchmark graphs carries over to a graph with over a hundred times the nodes of T-Finance: the tree ensembles and G-SNLS are the only methods near the ceiling, separated by at most 0.05% AUROC, and G-SNLS holds the best macro F1. For scale, GADBench reports on this graph, with its own search per method, 95.7 to 97.1% AUROC and 75 to 87% AUPRC for GraphSAGE, GCN, BWGNN and GHRN, 97.6 and 97.3% AUPRC for RF-Graph and XGB-Graph, and 99.9% AUROC with 99.1% AUPRC for BGNN, a gradient-boosting hybrid we do not run; our full-batch convolutions at their T-Finance configuration fall below those tuned values, and G-SNLS reaches its scores without any search on this graph. One G-SNLS trial on this graph takes between two and four hours on one GPU at width 64, with the Chebyshev realization and a constant-memory backward pass. DGraph-Fin is distributed only through a registration-gated download at dgraph.xinye.com (site is currently down), which we could not complete within the response period; we will add it as soon as the files are available to us.

Q1. AUPRC without the ranking decoder, and the strongest baselines under the same focal plus smooth-AP objective.

Without the ranking decoder. Table 9 (Appendix F) scores that ablation on test AUPRC at the epoch of best validation AUROC on the six ablation datasets. Removing the ranking decoder changes AUPRC by +0.24+0.24% on Minesweeper, −0.17-0.17% on Tolokers, −1.36-1.36% on Questions, −0.33-0.33% on Amazon, +0.04+0.04% on Elliptic and +0.39+0.39% on inj-Flickr; the paired test resolves only the inj-Flickr change, and in the direction of removal. The AUPRC margins over the strongest baseline (12.6% on Minesweeper, 7.4% on inj-Flickr, 6.1% on Tolokers) therefore do not come from the decoder that targets the metric. What the decoder contributes is the 0.41% of AUROC on Questions that the extended ablation reports (Table 10).

Baselines under our objective. Table 14 (Appendix G) retrains GraphSAGE, BWGNN, GHRN and SAGAD with their final linear layer replaced by our two decoders, one under the focal loss and the other under the smooth average-precision surrogate with the same constants, every other validated setting unchanged. On AUROC the objective changes a backbone by less than 0.7% in 22 of the 32 backbone-dataset pairs and lowers it by 1 to 5.5% in the other ten, nine of them on Questions and the two injected datasets; the largest gain is +0.62%. On AUPRC it lifts every backbone on T-Finance by 1.7 to 2.8% and GHRN on inj-Amazon by 7.6%, and lowers GraphSAGE and SAGAD on inj-Amazon by more than 10%. The retrained backbones stay below G-SNLS on every dataset except T-Finance, where BWGNN under our objective reaches 97.17% against our 97.01%. The objective is part of the method, but it accounts for a small part of the margin; the components the ablations measure account for the rest.

Dear Reviewer (3A4j), We would like to sincerely thank you for your valuable feedback and thoughtful consideration. If you feel that your concerns have been satisfactorily addressed, we would be truly grateful if you could kindly consider revising the score upward, as it would greatly improve the chances of our paper being accepted.

Reviewer_X82WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes G-SNLS, a model for supervised graph anomaly detection. It argues that some anomalies resemble their neighbors while others differ sharply from them, so a single spectral filter cannot handle both. G-SNLS splits the node features into heads, and each head applies a damped Schrödinger-type filter with four learned parameters for damping, dispersion and time scale. In each layer, a complex channel map then mixes the heads, and a nonlinear modulation adjusts the signal at each node. Two decoders, trained with a focal loss and an average-precision loss, produce the final scores. The method is evaluated on eight benchmarks against nine baselines, together with ablations and further analyses.

Critical strengths

S1. The spectral filter uses very few parameters, and each one has a clear meaning. A free polynomial filter learns K coefficients per head and per layer, while G-SNLS uses four numbers per head, shared across all layers. Damping sets how strongly each frequency is attenuated, and dispersion shifts its phase, so the filter is not limited to smoothing. The free-filter control shows that this compact form reaches about the same AUROC as a free filter of the same order. The method therefore gets a small, interpretable filter at little or no cost in accuracy.

S2. The experimental study is extensive. Besides the main comparison with nine baselines on eight benchmarks, the paper includes component ablations, a sensitivity study, an analysis of how the filter form affects optimization, and runtime and memory measurements. These analyses help show what each part of the model contributes.

Critical weaknesses

W1. The propagation layer appears to be an incremental extension of prior work that the paper does not discuss. Unitary graph convolutions (Kiani et al., 2024) already apply a complex exponential propagator followed by a learned complex channel matrix. FLODE (Maskey et al., 2023) learns a Schrödinger-type graph ODE with a complex channel matrix, which already gives each channel its own damping and dispersion. Compared with these, the main additions are the quadratic dispersion term and the nonlinear modulation. Discussing these methods would clarify what is new.

W2. The link between the proposed dynamics and anomaly detection is not clearly established. The paper motivates the design by arguing that camouflaged and clashing anomalies need different frequency responses at the same time, but it does not directly show this, and it is unclear why phase rotation in particular helps identify anomalies. The experiments mainly show that the best response differs across datasets, which learnable spectral filters are already designed to handle. Using several heads helps only on Minesweeper and Tolokers, by less than 1% AUROC. The largest gains come from the channel map and the modulation, which are not specific to anomaly detection. A clearer explanation of why this architecture suits anomaly detection would strengthen the paper.

W3. The ablations remove the channel map and the modulation entirely, which leaves a model with no channel mixing or no nonlinearity. This shows that these parts are needed, but not that their specific form is. A comparison with standard alternatives, such as a real linear layer and a common nonlinear activation like ReLU, would clarify this.

W4. The empirical comparison has two gaps. First, it omits strong baselines for supervised graph anomaly detection, such as XGB-Graph and RF-Graph from GADBench (Tang et al., 2023), PMP (Zhuo et al., 2024) and DSGAD (Zheng et al., 2025). Second, G-SNLS selects its architecture separately for each dataset, including the input context, readout, potential network, dispersion order and number of heads. Results with one shared architecture across datasets would show how much of the gain depends on this selection.

W5. The paper describes several baselines as fixing their frequency response before training, but some of them learn it. GPR-GNN learns its polynomial filter coefficients, and the paper's own related work says it "learns signed polynomial responses". AMNet also learns the coefficients of its Bernstein filters during training. The spectral parameters of G-SNLS are likewise learned during training and fixed afterwards. The difference is therefore in how the response is parameterized, not in whether it is learned. Stating this more precisely would make the contribution clearer.

Questions for authors

Check weakness

4 comments

We thank the reviewer for the detailed reading and for the two strengths noted. The five weaknesses asked for specific experiments and clarifications, and we ran or made all of them.

W1. Relation to unitary graph convolution (Kiani et al., 2024) and fLode (Maskey et al., 2023).

We agree these two methods should be discussed and compared, and we did both.

Discussion (Section 2). Unitary graph convolution applies exp⁡(iA)\exp(iA) through a truncated Taylor series followed by a learned complex channel matrix, and fLode takes explicit Euler steps of x′=−iLαxWx' = -iL^\alpha xW with a learned exponent, a learned step size and a diagonal complex channel matrix, so each fLode channel indeed carries a complex rate of its own. G-SNLS differs in three respects that we now state explicitly. (i) The propagator is the exact exponential of a generator whose damping γh\gamma_h and quadratic dispersion ωh(λ)\omega_h(\lambda) are learned per head, so attenuation and transport of a band are separate learned quantities and the composed response stays in the family at every depth (Proposition 2); a unitary layer preserves the norm by construction and cannot attenuate, and an Euler step with a per-channel complex rate has no such closure and can amplify a band. (ii) A deterministic intensity-dependent modulation sits between the steps; neither method has a nonlinearity between its propagation steps, and the dissipative form of ours gives the amplitude a ceiling (Proposition 6) and is the component whose removal costs most (Table 5). (iii) The readout and the two-decoder objective are built for scarce positives.

Comparison. We ran the released UniConv layer and the released fLode model under the same two-stage search as every baseline, and we also placed the unitary propagator inside our own architecture (undamped propagator with the ReLU of Kiani et al. in place of the modulation, and without any modulation) as the last two rows of Table 5. UniConv reaches 83.64, 72.31, 98.03, 97.04, 98.60, 92.09, 94.41 and 95.18% AUROC and fLode 83.40, 75.39, 97.85, 96.41, 98.03, 84.95, 91.83 and 93.38% on Tolokers, Questions, Amazon, T-Finance, Elliptic, inj-Amazon, Minesweeper and inj-Flickr, against 85.11, 76.78, 98.01, 97.01, 99.23, 90.18, 95.55 and 95.08% for G-SNLS. UniConv ties G-SNLS on Amazon, T-Finance and inj-Flickr, passes it by 1.9% on inj-Amazon and trails by 1.5% on Tolokers, 4.5% on Questions and 1.1% on Minesweeper; fLode trails on every dataset, by 0.2% on Amazon and up to 5.2% on inj-Amazon. On AUPRC G-SNLS is ahead of both on every dataset except T-Finance, where UniConv leads by 0.1%. One implementation note: the released UniConv layer forms one complex message per edge for every Taylor term, which does not fit Amazon or T-Finance; we apply the same normalized adjacency as a sparse product on the real and imaginary parts, which returns the same outputs and gradients (verified to float precision), and the appendix states this.

W2. Why these dynamics suit anomaly detection, and why phase rotation helps.

We take the three parts of the comment in turn.

Why phase rotation helps. A layer passes on the real part of the state, so the response a head applies to the next layer is e−Δtshγhλcos⁡[Δtshωh(λ)]e^{-\Delta t s_h\gamma_h\lambda}\cos[\Delta t s_h\omega_h(\lambda)] mixed with the corresponding sine term through the channel map. The cosine changes sign across the spectrum, so a head with dispersion is a band-pass or oscillatory filter with two parameters, whereas a head with damping alone is a monotone low-pass filter. This is exactly the shape that anomalies which clash with their neighborhood need (BWGNN's beta wavelets are band-pass for the same reason), and the shape that camouflaged anomalies do not need. The two regimes therefore sit in two parameters per head rather than in two architectures, and the ablations separate them: removing dispersion costs 3.86% on Minesweeper, whose mines live at high frequency, and nothing the paired test resolves on Questions, whose validated dynamics reduce to one damped head, while removing damping costs 4.97% on Questions and 1.49% on Tolokers and nothing resolved on Minesweeper (Table 5). The learned values follow the same split (Table 24 in Appendix L, first split): the four Minesweeper heads learn damping between 0.59 and 0.76 with quadratic dispersion up to 1.09 and scales spanning a factor ten, whereas the single Questions head learns damping 2.04 with a quadratic dispersion of −0.04-0.04, a strongly damped flow with almost no curvature in its phase.

Multiple heads. We agree that the head count is a minor axis, and the paper says so (Appendix J: four graphs are flat along it). The claim of the paper is not that many heads are needed but that damping and dispersion are learned rather than fixed, which the single-head configurations of Questions and inj-Flickr also use.

The channel map and the modulation are not specific to anomaly detection. The components are generic; their form is not, and W3 asked for exactly that test. The modulation is dissipative: its cubic term enters the modulus and caps the amplitude every layer emits (Proposition 6), and the AUPRC paragraph of Section 4.2 ties that ceiling to the metric where positives are scarce. Replacing it by a ReLU costs 1.71% on Minesweeper, 0.51% on Tolokers, 7.63% on Questions, 0.17% on Elliptic and 4.14% on inj-Flickr and changes Amazon by +0.13%, so what matters is the dissipative form, not the presence of a nonlinearity. The complex channel map is the one component whose form we cannot defend on these data: a real linear map changes AUROC by +0.83% on Minesweeper, −0.23-0.23% on Tolokers, +0.15% on Questions, +0.04% on Amazon, −0.08-0.08% on Elliptic and −0.10-0.10% on inj-Flickr. The mixing itself matters (removing it costs up to 6.72%, Table 5), its complex form does not. We report this plainly in the revision; Theorem 3 holds for either map because it enters only through its spectral norm.

W3. Standard alternatives instead of removal.

We completed this as well in the last four rows of Table 5 (blue): a real linear channel map, a ReLU in place of the modulation, and the unitary propagator with the ReLU or with no modulation. The real map stays within ±0.83\pm0.83% everywhere. The ReLU loses 1.71% on Minesweeper, 0.51% on Tolokers, 7.63% on Questions, 0.17% on Elliptic and 4.14% on inj-Flickr and gains 0.13% on Amazon. The unitary propagator with the ReLU loses 1.31%, 1.24%, 6.13%, 0.42% and 1.01% on the same five datasets and 0.02% on Amazon, and without any modulation it loses 3.70%, 2.21%, 6.71%, 0.46% (Amazon), 1.43% and 2.15%. Of the three choices the reviewer names, the dissipative modulation and the damped propagator carry the accuracy and the complex channel map does not.

W4. Missing baselines and per-dataset architecture selection.

Baselines. We added XGBGraph and RF-Graph (GADBench), PMP, and, since Reviewer 3A4j asked for them, ConsisGAD and SpaceGNN, together with MagNet, UniConv and fLode, on the same ten splits (Appendix G, Table 11). XGBGraph and RF-Graph receive the full two-stage search over the GADBench search space, PMP its released recipe with the learning rate chosen on validation, and ConsisGAD and SpaceGNN their released configurations. We could not find public code for DSGAD (Zheng et al., AAAI 2025), whose authors report four datasets with trainable wavelets; we now cite it in the related work and will add it if the code becomes available. Across the eight additions G-SNLS keeps the best AUROC on Tolokers and Minesweeper (resolved against all eight), is second on Questions (SpaceGNN, +0.36%) and on inj-Flickr (UniConv, +0.10%), and keeps the best AUPRC on those four datasets. XGBGraph leads on Amazon, T-Finance and Elliptic by 0.30 to 0.40% and on inj-Amazon by 8.7%, where SpaceGNN (95.13%), PMP (93.67%) and UniConv (92.09%) are also ahead of our 90.18%. The validated inj-Amazon configuration is also the one the restricted selections improve on (95.09% under the baseline budget), so part of that gap is selection rather than capacity; the tables keep the submitted number.

One shared architecture. Table 22 (Appendix K) reruns G-SNLS with a single configuration for every dataset, chosen as the most frequent validated value of each setting (width 128, eight steps of 0.75, four heads, raw input, quadratic dispersion, signed readout, linear potential, learning rate 3e-3, weight decay 1e-4, dropout 0.2, focal exponent 1) with no search at all, and under the baseline budget of eighteen sampled configurations. With the shared configuration G-SNLS reaches 85.34, 72.90, 97.57, 97.12, 99.03, 91.50, 94.89 and 94.44% AUROC on Tolokers, Questions, Amazon, T-Finance, Elliptic, inj-Amazon, Minesweeper and inj-Flickr, within one point of the validated selection on six datasets, 3.9% below on Questions and 1.3% above on inj-Amazon, and ranks 2.12 on AUROC and 1.50 on AUPRC among the nine baselines. Under the baseline budget it reaches 85.56, 75.25, 98.07, 97.03, 99.19, 95.09, 97.42 and 95.78 and keeps the validated ranks of 1.12 and 1.00. The per-dataset architecture selection is therefore not what produces the ranking.

W5. "Fixing the response before training."

We thank the reviewer for this comment. GPR-GNN learns its polynomial coefficients, AMNet learns the coefficients of its Bernstein pair, and our own parameters are likewise learned and then frozen. The difference is in how the response is parameterized and what the parameterization guarantees, not in whether it is learned. We corrected the three places that said otherwise (introduction, the AMNet sentence in Section 4.1 and the average-rank paragraph of Section 4.2, all blue). The free-response control of Section 4.4 is the comparison this distinction calls for: at matched polynomial order a learned free polynomial reaches the same accuracy, and what the constrained parameterization adds is four parameters per head shared across layers, a response that cannot leave the unit ball at any depth, and the flattest peak curvature per unit change in the response on all six datasets.

Reviewer_X82WOpenReview ↗

Maintain my score

Thank you for the detailed response and the additional experiments. However, the novelty remains incremental over unitary graph convolution and FLODE, and with the new baselines the empirical advantage is not consistent: G-SNLS has the best AUROC on only two of the eight datasets, and XGBGraph leads on four. I therefore keep my rating.

Reviewer_k2ijOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose G-SNLS, a new spectral graph neural network for anomaly detection. This new model is based on nonlinear Schrodinger dynamics, where the model splits its channels into spectral heads to incorporates nonlinear intensity-dependent absorption, node-specific potential, and channel mixing to capture complex graph patterns while maintaining bounded representations. The authors establish theoretical properties of G-SNLS and conduct experiments on eight anomaly detection benchmarks.

Critical strengths

The proposed model is novel. It is motivated by an interesting intuition from complex-valued dynamical systems. The authors further provide theoretical guarantees, showing how spectral damping controls energy decay while dispersion governs phase evolution, which strength the motivation behind the proposed model.

Critical weaknesses

The theoretical analysis is interesting, but guarantees mainly concern the linear spectral propagator rather than the complete NN architecture.

Although the empirical study is extensive, the performance gains are small. Several baselines achieve comparable results, where the margins of the proposed model over the strongest baselines are often small relative to the reported standard deviations.

Questions for authors

The use of complex-valued representations and phase information in GNNs is not entirely new, as similar ideas have been explored in prior work such as magnetic-Laplacian GNNs (Zhang et al., 2021; Fiorini et al., 2023). Can you include experimental comparisons with these methods to better demonstrate the advantages of G-SNLS?

4 comments

We thank the reviewer for the positive assessment of the model's novelty and of the theoretical motivation, and we address the three points below.

W1. The guarantees mainly concern the linear propagator.

Two of the six results concern the propagator alone, one the Chebyshev realization and three the complete architecture, and we agree the submission did not make this division explicit.

  • Theorem 1 (energy law) and Proposition 2 (closure under depth) characterize the response family of the propagator bank. They are stated on the propagator because that is the object the parameterization constrains: they hold at every point of training for every head, whatever the rest of the network does.
  • Theorem 3 (full-depth state bound) is the statement about the complete layer. Its proof composes the propagator with the complex channel map, the nonlinear modulation, the real projection and the radial clip in the order the layer applies them, and bounds the state at the readout by min⁡{BNd, [m(1+ε)]T∥ψ0∥F}\min\{B\sqrt{Nd},\,[m(1+\varepsilon)]^T\|\psi_0\|_F\} with mm the spectral norm of the channel map. Proposition 6 (absorption ceiling) is a property of the nonlinear modulation on its own: every entry it emits is at most (2e aμh)−1/2(2e\,a\mu_h)^{-1/2} for any input. Proposition 4 (permutation equivariance) covers the context encoder, the potential, the evolution, both readouts and both decoders.

We added a paragraph after Theorem 3 that says this (Section 3.3, blue), and a new Appendix L that measures the trained models against the bound. On the first split of every dataset the state norm stays within a factor 1.7 of the encoded state on six datasets, within 3.4 on inj-Flickr and grows to 42 over the eight layers on inj-Amazon, against bounds between 81 and 5×1075\times10^7; the radial clip touches no entry, and the largest entry leaving a modulation equals the ceiling of Proposition 6 to three decimals on every dataset (Table 23). The bound is loose because it compounds mm over depth, but what it guarantees, a finite state at any depth without the clip ever acting, is what the trained models show. The free-response control of Section 4.4 is the empirical counterpart: a free polynomial of order 32 collapses on three of ten inj-Flickr splits because nothing bounds its response, and the constrained response admits no such failure (Theorem 1 and Proposition 2).

W2. The gains are small relative to the reported standard deviations.

The standard deviation in the tables is the spread across the ten random splits, which every method shares. The question whether a lead is real is therefore a paired one, and the paper tests it that way: on the same splits, a paired Wilcoxon test with Holm correction across the nine baselines resolves twelve of the sixteen ranking comparisons at p<0.05p<0.05, and no baseline holds a resolved lead over G-SNLS anywhere. The average rank is 1.12 on AUROC and 1.00 on AUPRC among ten methods, and the Friedman test rejects rank equality on both. On average precision the margins over the strongest baseline are not small: 12.6% on Minesweeper, 7.4% on inj-Flickr, 6.1% on Tolokers and 2.3% on Elliptic. Against the eight baselines added at review time the picture is the same: the AUROC lead on Tolokers and Minesweeper is resolved against every one of them, the AUPRC lead holds on Tolokers, Questions, Minesweeper and inj-Flickr, and where a new baseline is ahead, XGBGraph on Amazon, T-Finance and Elliptic by 0.30 to 0.40% and on inj-Amazon by 8.7%, mentioned in Appendix G.

Q1. Comparison with magnetic-Laplacian networks (MagNet, SigMaNet).

We ran this comparison and added it to Appendix G (Table 11). One fact explains the outcome. The phase of the magnetic Laplacian is exp⁡(2πiq (A−A⊤))\exp(2\pi i q\,(A-A^\top)), so it encodes edge direction. Every benchmark in the paper (and in GADBench) is undirected, so that term is identically one for every charge qq, the imaginary part of the MagNet state stays zero at every layer, and MagNet reduces exactly to a real Chebyshev network with ReLU whose unwind layer reads the real part. The sign-magnetic Laplacian of SigMaNet reduces to the same operator on an undirected, unsigned graph. We therefore ran MagNet with its released architecture under the same two-stage search as every other baseline, which on these graphs is the Chebyshev network it reduces to, and one run covers both methods. MagNet reaches 93.81% AUROC on Minesweeper, 84.58% on Tolokers, 75.37% on Questions, 97.88% on Amazon, 97.04% on T-Finance, 98.77% on Elliptic, 89.48% on inj-Amazon and 94.86% on inj-Flickr, against 95.55, 85.11, 76.78, 98.01, 97.01, 99.23, 90.18 and 95.08 for G-SNLS: it trails on seven datasets and ties on T-Finance.

The complex state of G-SNLS comes from a different place. Its phase is the dispersion ωh(λ)\omega_h(\lambda) of the generator, a function of the Laplacian eigenvalue rather than of edge direction, so it is present on undirected graphs. It turns the real part carried across a layer boundary into a sign-changing response in λ\lambda with two parameters per head, and the "No dispersion" row of Table 5 measures its contribution directly: 3.86% AUROC on Minesweeper and 1.12% on Tolokers, with the paired test resolving both. We added this clarification to the related work.

Dear Reviewer (k2ij), We would like to sincerely thank you for your valuable feedback and thoughtful consideration. If you feel that your concerns have been satisfactorily addressed, we would be truly grateful if you could kindly consider revising the score upward, as it would greatly improve the chances of our paper being accepted.

Thank you for the detailed response. Some of my concerns are addressed; therefore, I raise the rating. I remain concerned that the empirical evidence is not substantially stronger than the baselines from the added experiment results in Appendix G.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal

We thank all three reviewers. This comment lists the summary of our rebuttal. Revised text is marked in blue in the PDF.

Reviewer k2ij

ConcernsComments
W1. Guarantees mainly concern the linear propagatorStated which results cover what: Theorem 1 and Proposition 2 concern the propagator; Theorem 3 and Propositions 4 and 6 cover the complete layer. Added a paragraph after Theorem 3 and Appendix L (Table 23), which measures trained models against the bound.
W2. Gains are small relative to standard deviationsClarified that the deviation is the spread across shared splits, so the comparison is paired. Wilcoxon with Holm correction resolves 12 of 16 comparisons against the nine original baselines, none of which holds a resolved lead.
Q1. Compare with magnetic-Laplacian GNNsAdded MagNet (Table 11). On undirected graphs MagNet and SigMaNet reduce to a real Chebyshev network; it trails on seven datasets and ties on T-Finance. Clarified in related work.

Reviewer X82W

ConcernsComments
W1. Relation to unitary graph convolution and FLODEAdded the discussion to Section 2, ran both as baselines (Table 11), and placed the unitary propagator inside our architecture (Table 5).
W2. Why these dynamics suit anomaly detectionExplained that dispersion yields a sign-changing band-pass response, supported by the dispersion and damping ablations (Table 5) and the learned parameters (Table 24).
W3. Standard alternatives instead of removalAdded four rows to Table 5: a real linear channel map, a ReLU, and the unitary propagator with and without modulation. The dissipative modulation and the damped propagator carry the accuracy; the complex form of the channel map does not, and the revision says so.
W4. Missing baselines; per-dataset architecture selectionAdded XGBGraph, RF-Graph and PMP (Table 11). DSGAD has no public code and is now cited. Added Table 22 with one shared configuration and with the baseline search budget.
W5. Baselines described as fixing their response before trainingCorrected the wording in the introduction and in Sections 4.1 and 4.2.

Reviewer 3A4j

ConcernsComments
W1. Is the contribution Schrödinger-specific?Clarified the claim: the constraint does not raise accuracy over a matched free filter. It gives fewer parameters, a bounded response and better conditioning at no cost in accuracy (Section 4.4, Table 17, Appendix I).
W2. Stability claim versus the full frameworkTheorem 3 covers the complete layer, and Proposition 2 is now stated as propagator-only. Appendix L reports the measured state norms.
W3, Q2. Tuning cost and equal search budgetsTable 22 reruns G-SNLS with 18 sampled configurations, which keeps the average ranks of 1.12 (AUROC) and 1.00 (AUPRC). Per-epoch cost is in Appendix M.
W4. More recent baselines and larger datasetsAdded XGBGraph, ConsisGAD and SpaceGNN (Table 11) and T-Social (Table 15). DGraph-Fin could not be downloaded within the response period.
Q1. AUPRC without the ranking decoder; baselines under our objectiveAdded Table 9 (decoder removed) and Table 14 (four baselines retrained with our two decoders).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.