{"id":"6e9ac369-dae9-4ed8-b91f-88946f041b20","arxiv_id":"2502.03201","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-space GNN with learnable curvature and distance-aware propagation that outperforms existing node anomaly detectors on 9 benchmark datasets under limited supervision.","lead":"SpaceGNN is a graph neural network that projects nodes into learnable Euclidean, hyperbolic, and spherical spaces to detect anomalous nodes in graphs with very few labeled examples. The paper reports large average gains over a recent pseudo-label baseline and provides code, but its headline comparison is over-stated in the abstract.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gain is computed against CONSISGAD on a single 50/50 random split with no variance, so the 'best rival' claim is not yet supported; a multi-seed rerun against the actual per-dataset best baseline is needed.","rationale":"The paper is transparent in providing code, ablations, and additional experiments, and I take the empirical engineering contribution seriously. However, the strongest_claim as phrased ('outperforms the best rival') is the load-bearing assertion, and it is the least secure part of the paper. The abstract's number is computed against CONSISGAD, not the best baseline, and Tables 1-2 contain no variance information. Section 5.1's split description is terse, but Appendix G's 10/10 and 100/100 tables show that the notation means 50 training nodes plus 50 validation nodes, so the main result rests on exactly one random draw of 50 labels. Given anomaly rates between 1% and 22%, such draws vary enormously and can even lack an anomalous training label on Reddit. The validation set is also only 50 nodes, so the grid search in Appendix D selects hyperparameters from a very noisy objective. None of this proves the method is bad; Appendix L's GADBench results with fixed 20-positive/80-negative labels and the training-size curves in Appendix K provide partial independent support. However, those experiments use a different protocol and do not pin down the headline 8.55%/4.31% under the 50/50 setting. A multi-seed rerun against the actual best per-dataset baseline is the single check that would settle the central claim. The reader's weakest_assumption (the Gaussian assumption in Theorem 1) is also fragile, and the proof's covariance simplification is questionable, but the central claim is empirical, so I place the experimental protocol concern first. The reader flagged the ambiguity and lack of error bars in the rationale, though not in the weakest_assumption field, so my agreement is partial.","tokens_in":29822,"tokens_out":5374,"duration_ms":52002,"concrete_test":"Recompute the headline comparison against the per-dataset best baseline for AUC and F1 separately, and rerun SpaceGNN plus the top 3 baselines (at least XGBGraph, CONSISGAD, and HGCN) on 20 random 50/50 train/validation splits, reporting mean plus/minus standard deviation and paired bootstrap p-values. If the average advantage over the per-dataset best rival drops below statistical significance on most datasets, or if SpaceGNN is not the best on Amazon F1 and other individual metrics across seeds, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim, stated in the abstract and Section 5.2, is that SpaceGNN 'outperforms the best rival by an average of 8.55% in AUC and 4.31% in F1 scores.' Two protocol facts make this claim insecure. First, the reported gain is the average advantage over CONSISGAD only, not over the best rival on each dataset. In Table 2, e.g., on Amazon, XGBGraph has F1 0.9231 while SpaceGNN has 0.8935, so SpaceGNN does not even beat the best rival on that metric; on Weibo AUC, XGBGraph (0.8421) is closer to SpaceGNN than CONSISGAD (0.7195). Second, Section 5.1 says datasets are 'randomly divided into 50/50 for training/validation,' and Appendix G confirms this means 50 labeled training nodes and 50 validation nodes. Tables 1-2 report one run with no standard deviation, no seed count, and no paired significance test. With only 50 random labels, split variance is large: for Reddit (3.33% anomaly rate), the expected number of anomalous training nodes is about 1.7, and many splits will contain zero anomalies, making the reported majority-class F1 values (0.4915) an artifact of the split rather than evidence of detection ability. The hyperparameters are also grid-searched on a 50-node validation set (Appendix D), compounding the noise. Appendix L's GADBench setting with fixed 20 positive and 80 negative labels is more controlled and provides genuine supporting evidence, but it uses a different protocol and does not validate the 8.55%/4.31% headline under the 50/50 protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SpaceGNN, a multi-space graph neural network for node anomaly detection (NAD) with extremely limited labels. It combines a learnable space projection (LSP) with learnable curvature, a distance-aware propagation (DAP) module motivated by a new weighted homogeneity metric, and a multiple-space ensemble (MulSE). The authors provide two theorems and two propositions as theoretical support, and report experiments on 9 real-world NAD datasets against 16 baselines, claiming average improvements of 8.55% in AUC and 4.31% in F1 over the best rival. The code is released.","tokens_in":30285,"tokens_out":10643,"duration_ms":89652,"significance":"If the empirical claims were fully supported, the paper would make a meaningful contribution: it offers a fresh geometric perspective on NAD under scarce supervision, introduces measurable concepts (expansion rate and weighted homogeneity), and provides an extensive comparison with code release. The controlled GADBench semi-supervised experiments (Appendix L) with fixed 20 positive and 80 negative labels give credible supporting evidence, and the ablation study shows that the LSP and DAP components each contribute. I do not see a circularity problem: the ER/WH analyses use labels for motivation, while the model is trained with cross-entropy on labeled nodes and tested on held-out nodes. However, the headline result is currently overstated because the reported margin is computed only against CONSISGAD, and the main 50/50 single-split protocol has large variance and can produce degenerate F1 scores on low-anomaly datasets. The paper's significance depends on whether these experimental weaknesses can be corrected.","major_comments":[{"comment":"The abstract and Section 5.2 state that SpaceGNN 'outperforms the best rival by an average of 8.55% in AUC and 4.31% in F1 scores,' but these figures are the average margins over CONSISGAD only, not over the best competing method on each dataset. In Table 2, XGBGraph achieves F1 0.9231 on Amazon while SpaceGNN achieves 0.8935, and CONSISGAD achieves F1 0.5084 on DGraph-Fin while SpaceGNN achieves 0.5017; the per-dataset best-baseline margins are therefore smaller, and the headline claim as written is not supported. Please restate the claim as a comparison with CONSISGAD or, preferably, report averages against the per-dataset best baseline.","section":"Abstract; Section 5.2, Tables 1-2"},{"comment":"Tables 1 and 2 report a single run on a single random 50/50 split with no standard deviation or number of seeds. Because the anomaly rates are very low (e.g., Reddit 3.33%, DGraph-Fin 1.27%), a training set of 50 nodes contains on average fewer than two anomalous nodes for Reddit, and many splits will contain no anomalous training nodes at all; the F1 value of 0.4915 shared by nearly all methods on Reddit is consistent with predicting the majority class. The single-split protocol therefore cannot support the claim of consistent superiority. Please run multiple random splits (or use stratified/fixed labels as in Appendix L), report mean±std, and include a paired significance test; Appendix L's fixed 20-positive/80-negative setting is a better-controlled protocol and should be promoted in the evaluation.","section":"Section 5.1; Tables 1-2; Appendix G"},{"comment":"The ensemble hyperparameters α and β chosen by validation (Table 4) collapse the model to a single space for 7 of the 9 datasets: for Weibo, T-Finance, DGraph-Fin and T-Social, β=1 so Z=Z_0 (Euclidean only), and for Reddit, Amazon and YelpChi, α=β=0 so Z=Z_{κ^-} (hyperbolic only). Only Tolokers and Questions actually use a mixture of spaces. Consequently, the main experiments do not demonstrate the benefit of the Multiple Space Ensemble module, and the ablation in Table 5 removes LSP and DAP but does not remove MulSE. Please add an ablation 'w/o MulSE' and either report results with the full ensemble active over all datasets or discuss why the validation procedure selects degenerate hyperparameters.","section":"Algorithm 4; Table 4; Table 5"},{"comment":"Hyperparameters (including α, β, learning rate, hidden dimension, layer count, and dropout) are selected by grid search to maximize the F1 score on a validation set of only 50 nodes. With such a small validation set, F1 differences of one or two points are within sampling noise, so the selection procedure is likely to overfit the validation split; this compounds the variance problem raised above. The authors should use a larger validation set, nested cross-validation, or fixed hyperparameters per protocol, and show that the reported test results are stable across choices.","section":"Appendix D; Section 5.1"}],"minor_comments":[{"comment":"The text says that when κ=0 the projection functions act as the identity, but substituting κ=0 into Equations (1)-(2) gives exp_o^0(x)=o (the origin) rather than x; the separate handling of κ=0 in Algorithm 1 should be stated in Section 3.","section":"Section 3, Eqs. (1)-(2)"},{"comment":"In the proof of Theorem 2, the Taylor expansion of tan^{-1}_κ(t) is written as t + κ t^3/3 in the κ→0+ and κ→0− cases, but then as t − κ t^3/3 in the concluding line; if the minus sign is intended, the expansion steps need to be reconciled, and the final formula in Theorem 2 should be re-verified.","section":"Appendix A, proof of Theorem 2"},{"comment":"The y-axis label '(Weighted) Hmogeneity' and the legend label 'Homogeniety' contain spelling errors; the figure would also benefit from a description of how the values are aggregated across nodes.","section":"Figure 3"},{"comment":"In the baseline list, 'AMNet' is spelled 'AMMNet'; please make the spelling consistent.","section":"Appendix B"},{"comment":"The statement 'randomly divide each dataset into 50/50 for training/validation' is ambiguous because it could be read as a 50% split; please clarify that it means 50 labeled nodes for training and 50 for validation.","section":"Section 5.1"},{"comment":"Theorem 1 assumes Σ_n ≈ Σ_a ≈ cI and independent Gaussian features; this assumption is not verified on the real datasets. Since the theorem is used to justify the distance-aware propagation, the authors should either provide empirical evidence for the assumption (e.g., per-class covariance estimates) or weaken the theoretical claim.","section":"Appendix A, Theorem 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an ICLR 2025 paper; for a journal submission, the experimental evaluation needs to be substantially strengthened before the central claim can be accepted. The architecture is plausible and the GADBench results provide partial support, so I recommend major revision rather than rejection. Please ensure that the 'best rival' phrasing is corrected and that the evaluation protocol is made robust to split variance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the architecture is real, the headline is misleading. If you take the abstract's 8.55% AUC claim at face value, you'll be over-reading what the tables actually show. The core method is sound and worth engaging with, but the empirical evidence needs another pass.\n\nWhat's new and good: this is the first application of a learnable-curvature multi-space GNN (Euclidean/hyperbolic/spherical ensemble) to supervised node anomaly detection with extremely limited labels. The Expansion Rate and Weighted Homogeneity analyses are genuinely new diagnostic tools, and the Distance Aware Propagation module is a reasonable attention mechanism. The argument that ensembling multiple spaces beats pseudo-label data augmentation, backed by the CONSISGAD false-rate analysis, is a useful contrast. Code is public, and the ablations confirm that both the learnable projection and the distance-aware propagation contribute.\n\nSoft spots, in proportion: the empirical foundation is thinner than the abstract implies. The 8.55% gain is against CONSISGAD, not the per-dataset best rival; Table 2 shows XGBGraph beating SpaceGNN on Amazon F1, and on several datasets the margin over the actual best baseline is much smaller. All main-table numbers come from one 50/50 random split with no standard deviation, no seed count, and no significance test. With 50 labeled nodes, split variance is high, especially on Reddit where the expected number of anomalous training nodes is about 1.7. Hyperparameters are grid-searched on the same 50-node validation set, compounding the selection noise. The theory is also conditional: Theorem 1 assumes independent Gaussian classes with nearly equal isotropic covariance, which is strong for real graph features, and Proposition 1 is essentially AM-GM. None of this is fatal. The GADBench semi-supervised setting with fixed 20 positive and 80 negative labels is more controlled and provides genuine supporting evidence, but it is a different protocol from the headline experiments.\n\nWho this is for: GNN researchers working on anomaly detection or scarce-label graph learning, and anyone comparing multi-space embeddings. This paper deserves a serious referee. If I were handling it, I'd ask for multi-seed reruns, error bars, and a corrected comparison against the best per-dataset baseline before trusting the magnitude of the gains. The ideas, however, are solid enough to build on.","headline":"A sensible multi-space GNN for scarce-label anomaly detection with a useful architectural story, but the headline claim of beating the best rival by 8.55% is computed against a single baseline on a single split and needs multi-seed verification.","tokens_in":30782,"tokens_out":2568,"would_cite":true,"duration_ms":23995,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SpaceGNN: multi-space graph learning detects anomalies with few labels","keywords":["node anomaly detection","graph neural networks","hyperbolic space","spherical space","limited supervision","learnable curvature","weighted homogeneity","model ensemble"],"falsifier":"Take a graph dataset with strongly anisotropic or non-Gaussian node features and compare SpaceGNN against a Euclidean-only variant with the distance-aware weighting removed. The proof of Theorem 1 predicts the ablation gap should track weighted homogeneity; if distance-aware propagation fails to help, or hurts, on such data despite high weighted homogeneity, then the Gaussian-covariance assumption is carrying the theoretical argument.","tokens_in":29629,"feed_emoji":"📈","tokens_out":3750,"duration_ms":35640,"temperature":0.7,"pith_summary":"The paper sets out to show that node anomaly detection under extremely limited supervision is better served by projecting graph data into several curved spaces at once than by staying in Euclidean space or by generating synthetic labels. It argues that different anomaly patterns, such as hierarchical structures, circular money-laundering rings, and plain relational data, have different natural curvatures, so a fixed Euclidean embedding misses information. The proposed SpaceGNN learns per-layer curvatures, weights message-passing edges by distance-based similarity, and ensembles one Euclidean, one hyperbolic, and one spherical GNN. Across nine real datasets, the authors report average gains of 8.55% in AUC and 4.31% in F1 over the strongest rival, with additional gains under a semi-supervised benchmark setting.","feed_headline":"Multi-space graph nets beat single-space rivals by 8.55% AUC","feed_subtitle":"With very few labels, ensembling Euclidean, hyperbolic, and spherical embeddings outperforms pseudo-label augmentation on 9 real datasets.","key_machinery":"The load-bearing machinery is the kappa-stereographic model of spaces of constant curvature, together with the expansion rate $ER_\\kappa = r_\\kappa/r_0$, which quantifies whether projecting a node triple into curvature $\\kappa$ enlarges the inter-class versus intra-class distance ratio. Weighted homogeneity $WH_\\kappa$ uses distance-based similarities $1-\\sigma(d_\\kappa(X_i, X_j))$ to weight same-label neighbor information and is shown, under Gaussian feature assumptions, to raise the probability that propagation preserves a node's class. These feed two modules: Distance Aware Propagation computes edge weights $\\omega^\\kappa_{ij} = \\text{MLP}(\\text{CONCAT}(X_i, \\hat{s}_{ij}X_j))$ using a Taylor-approximated distance, and Multiple Space Ensemble combines one Euclidean GNN with learnable-curvature hyperbolic and spherical GNNs. The theoretical anchors are Theorem 1, which says higher weighted homogeneity increases the chance that a node remains on its original distribution after propagation, and an ensemble bound showing that the ensemble cross-entropy loss is no worse than the weighted single-model loss.","core_discovery":"The central claim is that a multi-space ensemble with learnable curvature is a more effective response to scarce labels in node anomaly detection than pseudo-label generation or data augmentation. Concretely, the paper introduces an expansion rate measuring how a curved projection enlarges the ratio of inter-class to intra-class distance, uses it to motivate learnable curvature in a kappa-stereographic model, defines weighted homogeneity to justify distance-aware edge weighting during propagation, and proves an ensemble bound showing that combining models from different spaces reduces expected cross-entropy loss. On nine benchmark datasets, SpaceGNN is reported to outperform the best competing model by 8.55% AUC and 4.31% F1 on average.","pith_inferences":["A direct test of the expansion-rate rationale would regress the dataset-level performance gain of each curved branch against the empirical $ER_\\kappa$ computed from labeled triples; the paper reports learned curvatures but does not systematically make that connection.","The argument that pseudo-labeling is harmful rests on the observed false anomalous rate of one trained CONSISGAD model, so a natural extension is to vary pseudo-label confidence thresholds and see whether the ensemble advantage persists against calibrated pseudo-labeling.","Theorem 2 is a small-curvature expansion, so a stress test could rewire graphs to force large effective curvature and check whether Distance Aware Propagation degrades outside the approximation regime.","The framework treats anomaly detection as binary classification with imbalanced priors, leaving open whether the same space-ensemble idea transfers to unsupervised or single-class anomaly detection settings."],"forward_implications":["If the central claim is correct, scarce-label NAD systems can be improved without generating synthetic labels: an ensemble of models spanning several geometries substitutes for data augmentation.","Because the learned curvatures stay close to zero, the reported gains suggest that even modest curvature flexibility plus distance weighting drives most of the improvement.","The distance-aware propagation component gives a principled reason to down-weight noisy inter-class edges during message passing rather than only pruning them.","The ensemble bound implies that combining additional independent space-based models could further reduce expected loss, provided the base models remain sufficiently independent.","The reported semi-supervised results indicate the approach transfers to a standard 20-positive and 80-negative label regime on the same benchmark datasets."],"supporting_citations":[{"why":"Supplies the nine benchmark datasets, the GADBench evaluation protocol, and the XGBGraph baseline.","marker":"[Tang et al., 2023]"},{"why":"CONSISGAD is the recent pseudo-label baseline that SpaceGNN is designed to beat and whose generated labels are analyzed.","marker":"[Chen et al., 2024]"},{"why":"Provides the kappa-stereographic model and the projection functions used to build GNNs in spaces of different curvature.","marker":"[Bachmann et al., 2020]"},{"why":"HGCN supplies the non-Euclidean GNN formulation that motivates multi-space propagation.","marker":"[Chami et al., 2019]"},{"why":"GCN is the canonical Euclidean GNN baseline and defines the propagation framework being extended.","marker":"[Kipf & Welling, 2017]"},{"why":"Frechet inception distance is used inside the proof of Theorem 1 to measure distribution shift after propagation.","marker":"[Heusel et al., 2017]"},{"why":"Supports the claim that noisy pseudo-labels degrade performance, a premise for preferring model ensembling over pseudo-label augmentation.","marker":"[Wang et al., 2023a]"},{"why":"Used in the proof of Proposition 2 for the bias-variance decomposition of ensemble diversity.","marker":"[Wood et al., 2023]"}],"fun_headline_variants":["Multi-space GNN detects node anomalies with few labels","SpaceGNN: multi-space graph net lifts anomaly AUC by 8.55%","Scarce labels? Multi-space graph embedding beats augmentation","Ensembling Euclidean and curved spaces wins node anomaly detection","Learned curvatures give 8.55% AUC gain in node anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that weighted homogeneity helps propagation assumes normal and anomalous node features follow independent Gaussian distributions with roughly equal, near-scalar covariance $\\Sigma_n \\approx \\Sigma_a \\approx cI$; if real features are non-Gaussian or strongly anisotropic, that justification for distance-aware propagation is not established.","fun_headline_variants_meta":{"raw":{"variants":["Multi-space GNN detects node anomalies with few labels","SpaceGNN: multi-space graph net lifts anomaly AUC by 8.55%","Scarce labels? Multi-space graph embedding beats augmentation","Ensembling Euclidean and curved spaces wins node anomaly detection","Learned curvatures give 8.55% AUC gain in node anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00064,"raw_usage":{"total_tokens":2956,"prompt_tokens":961,"completion_tokens":1995,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":1918}},"tokens_in":577,"tokens_out":1995,"duration_ms":12441,"temperature":1.0,"reasoning_tokens":1918,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:32:32.529523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a graph dataset with strongly anisotropic or non-Gaussian node features and compare SpaceGNN against a Euclidean-only variant with the distance-aware weighting removed. The proof of Theorem 1 predicts the ablation gap should track weighted homogeneity; if distance-aware propagation fails to help, or hurts, on such data despite high weighted homogeneity, then the Gaussian-covariance assumption is carrying the theoretical argument.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nine benchmark datasets, the GADBench evaluation protocol, and the XGBGraph baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used in the proof of Proposition 2 for the bias-variance decomposition of ensemble diversity."}],"review_version":1}