Pith. sign in

REVIEW 5 major objections 5 minor 56 references

CurvGAD: Leveraging Curvature for Enhanced Graph Anomaly Detection

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read The paper claims that Ollivier-Ricci edge curvature, reconstructed by a mixed-curvature autoencoder, reveals graph anomalies that structure and attribute reconstruction miss, improving AUROC by up to 6.5%.

desk verdict Worth a serious look for the geometric decomposition idea, but the headline AUROC gains look like best-of-eight signature selection on the test set unless the authors can show a pre-registered or validation-based protocol. read the letter →

arxiv 2502.08605 v2 pith:LO5TVWS5 submitted 2025-02-12 cs.LG cs.AI

classification cs.LGcs.AI
keywords graphanomalydetectionOllivier-Riccicurvaturemixed-curvaturemanifoldsRicciflowautoencoderheterophilicgraphsgeometricanomaliesreconstruction-based
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a graph's intrinsic curvature, measured edge-by-edge as Ollivier-Ricci curvature, carries anomaly signal that conventional structure- and attribute-based detectors miss. It introduces CurvGAD, a mixed-curvature graph autoencoder with two parallel pipelines: one reconstructs the edge-curvature matrix from Riemannian node embeddings to expose geometric anomalies, and the other first flattens the graph with discrete Ollivier-Ricci flow and then reconstructs adjacency and features, isolating non-geometric anomalies. The claim is that this decomposition is both more accurate and more interpretable: on 10 real-world homophilic and heterophilic benchmarks, CurvGAD reports AUROC improvements of up to 6.5% over previous state-of-the-art methods, with the largest gains on heterophilic networks.

What carries the argument

The load-bearing object is the Ollivier-Ricci curvature (ORC) of an edge, defined as $\tilde{\kappa}_{xy} = 1 - W_1(m_x^\delta, m_y^\delta)/d_G(x,y)$, the Wasserstein-1 distance between lazy random-walk measures on the two neighborhoods, normalized by graph distance; the paper computes it with a linear-time combinatorial approximation from node degrees and triangle counts. The supporting machinery is a product manifold built from $\kappa$-stereographic hyperbolic, spherical, and Euclidean components; a filter bank of manifold-valued Chebyshev spectral filters acting on the discrete Laplace-Beltrami operator; a Gaussian-kernel curvature decoder $\tilde{C}_{xy} = 2\,\mathrm{sigmoid}(1 - \exp(-\gamma D_{\mathcal{M}}(n_x,n_y)^2/\tau^2)) - 1$; and discrete Ollivier-Ricci flow, which updates edge weights multiplicatively as $w^{(t+1)}_{xy} = (1 - \epsilon \tilde{\kappa}^{(t)}_{xy}) w^{(t)}_{xy}$ until curvatures converge toward zero. The curvature decoder is what turns geometry into an anomaly score, and the Ricci flow is what makes the second pipeline curvature-invariant.

What would settle it

On each of the 10 datasets, compute exact or high-precision ORC for every edge and report the share of edges with curvature outside $[-1,1]$; if a substantial share is outside, the curvature loss is dominated by clamping artifacts. A complementary experiment replaces the sigmoid-bounded curvature decoder with an unbounded linear decoder and checks whether AUROC and the set of top-ranked anomalies change materially.

Watch

Extended reading notes

Core claim

The central discovery the paper argues for is a new class of graph anomalies defined by curvature: edges or nodes whose Ollivier-Ricci curvature deviates sharply from the graph's expected geometric pattern, such as bottlenecks and hubs, which are invisible to pure adjacency or feature reconstruction. CurvGAD detects them by treating curvature reconstruction itself as a learning task: a product-manifold encoder with a Chebyshev filter bank maps nodes into hyperbolic, spherical, and Euclidean components, and a Gaussian-kernel decoder predicts edge curvature; nodes whose curvature is hard to reconstruct receive high geometric anomaly scores. A second, curvature-invariant pipeline deforms the graph under discrete Ollivier-Ricci flow until edge curvatures near zero, then reconstructs the adjacency and feature matrices in Euclidean space, so structural and attribute anomalies are scored separately from geometric ones. Combined, the two pipelines refine existing anomaly labels and surface curvature-driven outliers, and the paper's experiments report gains of 1.05% to 6.5% AUROC over the best baselines across the ten datasets.

Load-bearing premise

The load-bearing untested premise is that Ollivier-Ricci curvatures in the benchmark graphs fall within the decoder's output range $[-1,1]$, so high curvature-reconstruction error means geometric irregularity rather than decoder clipping.

Editorial extensions

If this is right

  • Reconstruction-based GAD detectors should add a curvature-reconstruction head, since curvature deviations are independent signal beyond adjacency and feature reconstruction.
  • Anomaly scores become decomposable: a node can be flagged as geometric, structural, attribute, or a combination, which makes detector output explainable in high-stakes settings.
  • The mixed-curvature filter bank gives a principled way to handle heterophilic graphs, where standard low-pass GNN assumptions fail; the largest reported gains are on heterophilic datasets.
  • Because ORC is approximated in linear time and Ricci flow converges in a small constant number of iterations, the curvature machinery scales to million-node graphs, as demonstrated on the largest datasets in the study.
  • If curvature anomalies are truly independent of labels, the pipeline can surface new, previously unlabeled outliers; at minimum it re-ranks known anomalies using geometric evidence.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to drop the sigmoid bound on the curvature decoder and use an unbounded output; if performance holds, the detected geometric anomalies are genuine curvature deviations, and if it drops, the reported scores partly measure decoder saturation.
  • The two-pipeline decomposition suggests a temporal variant: tracking Ollivier-Ricci curvature over graph snapshots could flag emerging bottlenecks or cascades before they become labeled anomalies, since Ricci flow already identifies the most distorted edges.
  • The same curvature-equivariant/invariant split could be carried over to edge-level and graph-level anomaly detection, and to directed graphs via directed curvature notions; the paper only evaluates node-level undirected graphs.
  • Interpreting 'curvature-driven anomalies' as ground-truth discoveries is stronger than the evidence: the AUROC numbers use existing labels, so a dedicated study would need to validate that high-curvature-error nodes are anomalous by independent criteria.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes CurvGAD, a dual-pipeline graph autoencoder for node-level anomaly detection. The first pipeline embeds nodes in a product of constant-curvature manifolds using a Chebyshev filter bank and reconstructs Ollivier-Ricci edge curvatures with a Gaussian-kernel decoder. The second pipeline regularizes the graph under discrete Ollivier-Ricci flow and reconstructs the adjacency and feature matrices in Euclidean space, so that structure and attribute reconstruction are decoupled from curvature. Anomaly scores combine the three reconstruction errors with a supervised classification loss. Experiments on 10 datasets report AUROC improvements of up to 6.5% over existing GAD baselines, with ablations over pipeline components and product-manifold signatures.

Significance. If the central claims hold, CurvGAD would be a useful contribution to graph anomaly detection: it introduces a curvature-based notion of geometric anomalies, offers an interpretable decomposition of anomaly sources, and reports strong AUROC results across both homophilic and heterophilic graphs. The paper has concrete strengths: the code is released, the ORC approximation is accompanied by a linear-time complexity argument, and the ablation study is fairly extensive. However, the empirical headline currently rests on a per-dataset model-selection protocol that is not fully specified, and on an unverified assumption about the range of the curvature decoder; both issues must be resolved before the reported gains can be attributed to the method as a single, reproducible model.

major comments (5)
  1. [§5.3, Table 1 vs. Table 3] The final CurvGAD AUROC for each dataset in Table 1 coincides exactly with the best row among the eight signatures in Table 3 (e.g., Reddit 70.42 = H24×S24, Amazon 99.62 = (H8)2×(S8)2×E16, YelpChi 89.33 = H8×S8×E32, Elliptic 90.13 = H16×S16×E16). The text says the signature is decided heuristically via Algorithm 2, but it never states that the signature was selected on a validation split before computing test AUROC. With 8 signatures and 10 datasets, selecting the best per-dataset result using test labels would inflate the reported gains even under a null model. The authors must specify the exact selection protocol and report the performance of the model obtained with validation-based signature selection, not the best of the ablated configurations.
  2. [§4.1.1, Eq. (4)] Equation (4) uses the symbol eLP, but eLP is never defined in the notation table or in the surrounding text. If eLP is intended to be the Laplace-Beltrami operator LP, the Möbius left-matrix-multiplication of LP with a manifold-valued embedding is not defined anywhere in the κ-stereographic model description of Appendix B.2. As written, the Chebyshev recursion in Eq. (4) is not reproducible and should be corrected with a precise definition of the operator and its action on manifold-valued vectors.
  3. [Definition 4.1 and Appendix D.1] The curvature decoder maps predicted curvatures into [-1, 1] via a sigmoid transformation, and Appendix D.1 justifies this only by asserting that it is 'consistent with the properties of Ollivier-Ricci curvature.' The paper reports no empirical distribution of ORC values on any dataset, and it does not quantify what fraction of true ORC values (or their Jost-Liu approximations) fall outside [-1, 1]. If a substantial fraction of edges have ORC outside this range, the Frobenius loss LC would be dominated by clamping artifacts and the geometric anomaly score would measure decoder saturation rather than curvature deviation. The authors should report the empirical ORC range for each dataset and verify that the decoder's output range matches the target values.
  4. [Appendix F, Algorithm 2; §4.2.1] Algorithm 2 chooses the product-manifold signature from the distribution of the same Ollivier-Ricci curvatures that are later used as the reconstruction target of the curvature loss LC, creating a selection dependence that is not acknowledged. In addition, Appendix F states that component dimensions are treated as tunable hyperparameters, and Appendix E.2 lists per-dataset tuning of δ, dP, F, and other settings. The evaluated system is therefore a family of per-dataset configurations rather than a single pre-specified model. The authors should either fix these choices a priori or describe a nested validation procedure; otherwise the reported up-to-6.5% improvement cannot be attributed to a single identifiable method.
  5. [Abstract, §1, Figure 1] The abstract and Section 1 claim that CurvGAD 'identifies new curvature-driven anomalies,' and Figure 1 highlights examples of such anomalies. However, no quantitative evaluation of these newly discovered anomalies is provided; the AUROC experiments are computed against existing labels only. The claim that genuinely new anomalies are detected is therefore unsupported. The authors should either temper this claim or provide additional validation, such as human annotation, known ground truth for a subset of candidate anomalies, or a case study with external evidence.
minor comments (5)
  1. [Appendix C.1.2] The phrase 'combinitorial bounds' should be 'combinatorial bounds,' and the same typo appears in the description of the Jost-Liu approximation.
  2. [Appendix D title] The appendix title 'Arcitectural Details' should be 'Architectural Details.'
  3. [Algorithm 2] Algorithm 2 contains the placeholder '[citation-withheld]' and states that the algorithm is adopted from an unpublished previous work. This prevents the reader from reproducing the signature-selection step; a proper citation or a full self-contained description is needed.
  4. [Table 5 / §C.2] The text in Appendix C.2 refers to 'Table C.2,' but the runtime table is numbered Table 5; the cross-reference should be corrected.
  5. [Section 5.5(c)] The ablation text says 'H24×S24 (E removed) and S24×E24 (H removed),' but the first configuration contains no Euclidean component by construction; this wording is misleading and should be clarified to say that the configuration omits a Euclidean component.

Circularity Check

2 steps flagged · score 4.0 of 10

Reported CurvGAD gains are the per-dataset best of an 8-signature ablation presented as a fixed model, with the signature heuristic backed only by an unpublished self-citation.

  1. fitted input called prediction [Section 5.3, Table 1 vs Section 5.5, Table 3]
    "The manifold signature is decided heuristically using the ORC distribution of the datasets (Algorithm 2, Appendix F). Table 3. Ablation Study on Product Manifold Signatures – AUCROC Score (Mean ± 95% CI). Best performing signatures are in Bold. CurvGAD 70.42±1.03 99.04±0.34 99.62±0.17 89.33±1.44 98.13±1.34 90.13±0.99 85.22±0.11 74.45±0.04 83.77±2.44 89.23±3.89"

    For every dataset, the final CurvGAD row of Table 1 reproduces the bolded best-signature row of Table 3: e.g., Reddit 70.42 is the H24×S24 entry, Amazon 99.62 is the (H8)2×(S8)2×E16 entry, Elliptic 90.13 is the H16×S16×E16 entry, DGraph 83.77 is the (H16)2×E16 entry, Questions 74.45 is the H24×E24 entry, and Tolokers 85.22 is the S24×E24 entry. Section 5.3 only states that the signature is decided heuristically using the ORC distribution and that ten random splits were averaged; no validation-set selection among the eight signatures is described. The reported up-to-6.5% gain is therefore the per-dataset maximum over eight configurations rather than the performance of one pre-specified model.

  2. ansatz smuggled in via citation [Appendix F, paragraph preceding Algorithm 2]
    "[citation-withheld] is used to respect the double blind policy for one of our previous works, which is still under review at the time of submission. Appropriate citation shall be inserted later. This algorithm has been adopted from our previous work."

    Algorithm 2 is the mechanism that fixes the product-manifold signature, the key dataset-dependent design choice on which the Table 1 results depend. The paper's only cited authority for this algorithm is the authors' own unpublished prior work, with the citation explicitly withheld. Since the algorithm is presented as a heuristic ('we do not claim that this heuristic yields the absolute optimal manifold decomposition'), the self-citation does not supply independent support; it imports an ansatz from unpublished same-author material and prevents the reader from checking whether the heuristic was derived before or after seeing the Table 3 AUROC values. The algorithm is, however, specified in the paper, so this is a supporting circularity rather than the sole basis of the results.

full rationale

The CurvGAD architecture itself is not definitionally circular: the curvature reconstruction target (ORC), the adjacency/feature decoders, and the supervised classification loss are distinct inputs, and the model is evaluated against external organic anomaly labels on ten datasets. However, the empirical headline is compromised by signature selection: the final numbers coincide exactly with the best row of the eight-signature ablation, and no validation-based selection protocol is reported, so the claimed improvement is partly a selected maximum rather than an independent prediction. The signature-estimation algorithm is additionally supported only by an unpublished same-author citation. Weighing these, the central architecture still has independent content, but the flagship empirical claim is partially forced by selection; score 4 rather than 0-2 because the reported AUROC reduction is concrete and quotable, though not a definitional equivalence.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on the ORC approximation, Ricci flow convergence, the heuristic curvature decoder, the undirected graph assumption, and per-dataset geometric hyperparameters. The proposed 'curvature-based geometric anomaly' entity is introduced without an external falsifiable benchmark.

free parameters (6)
  • Per-dataset product manifold signature (component types, dimensions, initial curvatures) = e.g., Reddit H24 x S24; Amazon (H8)2 x (S8)2 x E16 (Table 3)
    The choice of hyperbolic/spherical/Euclidean components and their dimensions changes the embedding geometry, directly affecting curvature reconstruction and anomaly scores; dimensions are treated as tunable hyperparameters per dataset (Appendix F).
  • Ollivier-Ricci lazy walk parameter delta = 0.5
    Controls the probability mass split between a node and its neighbors in ORC; chosen by hand as 'equal division' without a data-driven justification (Appendix B.3).
  • Gaussian kernel width gamma = not reported
    Fixed kernel width in the curvature decoder; no value or sensitivity analysis in the main text.
  • Chebyshev filter count F = 8
    Number of spectral filters in the filter bank; tuned per dataset (Appendix E.2).
  • Product manifold total dimension dP = 48
    Total embedding dimension; tuned per dataset (Appendix E.2).
  • Loss trade-off parameters lambda_C, lambda_A, lambda_X, lambda_cls = learned during training
    Learnable weights balancing reconstruction and classification losses; fitted to the labeled data.
assumptions (5)
  • domain assumption Ollivier-Ricci curvature approximated by Jost-Liu combinatorial bounds is accurate enough for anomaly detection.
    The linear-time approximation (Appendix C.1.2) replaces exact ORC; if approximation error is large, the curvature reconstruction target and the geometric anomaly scores are distorted.
  • domain assumption Ollivier-Ricci flow converges to near-zero curvature within 12-13 iterations on the benchmark graphs.
    Section 4.2.1 and Appendix C.1.3 rely on this empirical convergence to justify replacing the product manifold with a Euclidean manifold for structure/attribute reconstruction.
  • ad hoc to paper A Gaussian kernel on geodesic distances in the product manifold can model edge ORC values.
    Definition 4.1 and Appendix D.1 provide intuition but no derivation connecting ORC to embedding distances; the sigmoid restricts outputs to [-1,1], which may not match true ORC range.
  • domain assumption The input graph is undirected for ORC and Ricci flow computations.
    Algorithm 1 requires an undirected graph, but datasets like T-Finance and Elliptic are directed transaction networks; no symmetrization or directed ORC variant is described.
  • standard math The cotangent discretization of the Laplace-Beltrami operator is a valid graph Laplacian for spectral filtering in mixed-curvature spaces.
    Section 3 invokes Belkin et al. and Crane; assumed as background for the Chebyshev filter bank.
invented entities (1)
  • Curvature-based geometric anomalies
    purpose: A new anomaly class defined by deviations in task-specific Ollivier-Ricci curvature patterns, targeted by the curvature-equivariant pipeline.
    No ground-truth labels for such anomalies are used; Figure 1 is qualitative, and AUROC is computed only on existing structural/attribute anomaly labels, so the existence and detectability of this new class is not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CurvGAD: Leveraging Curvature for Enhanced Graph Anomaly Detection." pith.science (2026). https://pith.science/paper/LO5TVWS5

@misc{pith2026250208605,
  author       = {Pith},
  title        = {Pith review of: CurvGAD: Leveraging Curvature for Enhanced Graph Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LO5TVWS5}},
  note         = {Machine review of arXiv:2502.08605}
}
read the original abstract

Does the intrinsic curvature of complex networks hold the key to unveiling graph anomalies that conventional approaches overlook? Reconstruction-based graph anomaly detection (GAD) methods overlook such geometric outliers, focusing only on structural and attribute-level anomalies. To this end, we propose CurvGAD - a mixed-curvature graph autoencoder that introduces the notion of curvature-based geometric anomalies. CurvGAD introduces two parallel pipelines for enhanced anomaly interpretability: (1) Curvature-equivariant geometry reconstruction, which focuses exclusively on reconstructing the edge curvatures using a mixed-curvature, Riemannian encoder and Gaussian kernel-based decoder; and (2) Curvature-invariant structure and attribute reconstruction, which decouples structural and attribute anomalies from geometric irregularities by regularizing graph curvature under discrete Ollivier-Ricci flow, thereby isolating the non-geometric anomalies. By leveraging curvature, CurvGAD refines the existing anomaly classifications and identifies new curvature-driven anomalies. Extensive experimentation over 10 real-world datasets (both homophilic and heterophilic) demonstrates an improvement of up to 6.5% over state-of-the-art GAD methods. The code is available at: https://github.com/karish-grover/curvgad.

Figures

Figures reproduced from arXiv: 2502.08605 by the authors.

Figure 1
Figure 1. Curvature matters. The effect of curvature information on node-level anomaly scores for CurvGAD across Weibo, Reddit, Cornell and Chameleon datasets (Zhao et al., 2020). The plots show anomaly score distributions against curvature values, where the blue region highlights potentially new anomalies due to curvature deviations, and the red region highlights the effect of curvature on existing outliers. W/o Curv refers … view at source ↗
Figure 2
Figure 2. Architecture of CurvGAD. The proposed model employs two synergistic pipelines for anomaly detection: (1) Curvature￾Equivariant Reconstruction – embedding the input graph into a mixed-curvature product manifold using a spectral Chebyshev filter bank, with a Gaussian kernel-based decoder reconstructing the curvature matrix to capture curvature irregularities; (2) Curvature￾Invariant Reconstruction – regularizing the g… view at source ↗
Figure 3
Figure 3. Intuition on Ollivier-Ricci flow. Curvature regular￾ization of Karate Club graph (Rozemberczki et al., 2020) under Ollivier-Ricci flow. Observe how the flow increases (decreases) the weight of negatively (positively) curved edges. Red nodes indi￾cate curvature-based outliers near edges with extreme curvatures. Algorithm 1 Discrete Ollivier-Ricci Flow Require: An undirected graph G = (V, E), a small thresh￾old ∆ > 0,… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Runtime comparision of CurvGAD. Average runtime per epoch (training + inference) in seconds (log scale) on Reddit (Liu et al., 2022), Weibo (Zhao et al., 2020), Tolokers (Platonov et al., 2023), and Questions (Platonov et al., 2023) datasets, for select baselines. Data…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 49 canonical work pages

  1. [1]

    Constant curvature graph convolutional networks

    Bachmann, G., B \'e cigneul, G., and Ganea, O. Constant curvature graph convolutional networks. In International conference on machine learning, pp.\ 486--496. PMLR, 2020

  2. [2]

    Discrete laplace operator on meshed surfaces

    Belkin, M., Sun, J., and Wang, Y. Discrete laplace operator on meshed surfaces. In Proceedings of the twenty-fourth annual symposium on Computational geometry, pp.\ 278--287, 2008

  3. [3]

    Hyperbolic graph convolutional neural networks

    Chami, I., Ying, Z., R \'e , C., and Leskovec, J. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems, 32, 2019

  4. [4]

    Detecting network anomalies using forman--ricci curvature and a case study for human brain networks

    Chatterjee, T., Albert, R., Thapliyal, S., Azarhooshang, N., and DasGupta, B. Detecting network anomalies using forman--ricci curvature and a case study for human brain networks. Scientific reports, 11 0 (1): 0 8121, 2021

  5. [5]

    Hamilton’s Ricci flow, volume 77

    Chow, B., Lu, P., and Ni, L. Hamilton’s Ricci flow, volume 77. American Mathematical Society, Science Press, 2023

  6. [6]

    The n-dimensional cotangent formula

    Crane, K. The n-dimensional cotangent formula. Online note. URL: https://www. cs. cmu. edu/\ kmcrane/Projects/Other/nDCotanFormula. pdf , pp.\ 11--32, 2019

  7. [7]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Defferrard, M., Bresson, X., and Vandergheynst, P. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016

  8. [8]

    Deep anomaly detection on attributed networks

    Ding, K., Li, J., Bhanushali, R., and Liu, H. Deep anomaly detection on attributed networks. In Proceedings of the 2019 SIAM international conference on data mining, pp.\ 594--602. SIAM, 2019

Show all 56 references
  1. [9]

    Do Carmo, M. P. and Flaherty Francis, J. Riemannian geometry, volume 2. Springer, 1992

  2. [10]

    Dou, Y., Liu, Z., Sun, L., Deng, Y., Peng, H., and Yu, P. S. Enhancing graph neural network-based fraud detectors against camouflaged fraudsters. In CIKM, pp.\ 315--324, 2020

  3. [11]

    Cascade-lstm: A tree-structured neural classifier for detecting misinformation cascades

    Ducci, F., Kraus, M., and Feuerriegel, S. Cascade-lstm: A tree-structured neural classifier for detecting misinformation cascades. In proceedings of the 26th ACM SIGKDD international conference on Knowledge Discovery & Data Mining, pp.\ 2666--2676, 2020

  4. [12]

    Anomalydae: Dual autoencoder for anomaly detection on attributed networks

    Fan, H., Zhang, F., and Li, Z. Anomalydae: Dual autoencoder for anomaly detection on attributed networks. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.\ 5685--5689. IEEE, 2020

  5. [13]

    Learning mixed-curvature representations in products of model spaces

    Gu, A., Sala, F., Gunel, B., and R \'e , C. Learning mixed-curvature representations in products of model spaces. In International conference on learning representations, volume 5, 2019

  6. [14]

    L., Ying, R., and Leskovec, J

    Hamilton, W. L., Ying, R., and Leskovec, J. Inductive representation learning on large graphs. In NeurIPS, pp.\ 1025--1035, 2017

  7. [15]

    ADB ench: Anomaly detection benchmark

    Han, S., Hu, X., Huang, H., Jiang, M., and Zhao, Y. ADB ench: Anomaly detection benchmark. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  8. [16]

    Analyzing the effectiveness of graph metrics for anomaly detection in online social networks

    Hassanzadeh, R., Nayak, R., and Stebila, D. Analyzing the effectiveness of graph metrics for anomaly detection in online social networks. In Web Information Systems Engineering-WISE 2012: 13th International Conference, Paphos, Cyprus, November 28-30, 2012. Proceedings 13, pp.\...

  9. [17]

    Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection

    He, J., Xu, Q., Jiang, Y., Wang, Z., and Huang, Q. Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection. arXiv preprint arXiv:2312.14535, 2023

  10. [18]

    Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection

    He, J., Xu, Q., Jiang, Y., Wang, Z., and Huang, Q. Ada-gad: Anomaly-denoised autoencoders for graph anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 8481--8489, 2024

  11. [19]

    Bernnet: Learning arbitrary graph spectral filters via bernstein approximation

    He, M., Wei, Z., Huang, Z., and Xu, H. Bernnet: Learning arbitrary graph spectral filters via bernstein approximation. NeurIPS, 2021

  12. [20]

    Graph convolutional networks for traffic anomaly

    Hu, Y., Qu, A., and Work, D. Graph convolutional networks for traffic anomaly. arXiv preprint arXiv:2012.13637, 2020

  13. [21]

    Dgraph: A large-scale financial dataset for graph anomaly detection

    Huang, X., Yang, Y., Wang, Y., Wang, C., Zhang, Z., Xu, J., Chen, L., and Vazirgiannis, M. Dgraph: A large-scale financial dataset for graph anomaly detection. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

  14. [22]

    and Liu, S

    Jost, J. and Liu, S. Ollivier’s R icci curvature, local clustering and curvature-dimension inequalities on graphs. Discrete & Computational Geometry, 51 0 (2): 0 300--322, 2014

  15. [23]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In Proceedings of the ICLR, 2017

  16. [24]

    The hungarian method for the assignment problem

    Kuhn, H. The hungarian method for the assignment problem. Naval research logistics quarterly, 2 0 (1-2): 0 83--97, 1955

  17. [25]

    Predicting dynamic embedding trajectory in temporal interaction networks

    Kumar, S., Zhang, X., and Leskovec, J. Predicting dynamic embedding trajectory in temporal interaction networks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 1269--1278, 2019

  18. [26]

    Ricci curvature of graphs

    Lin, Y., Lu, L., and Yau, S. Ricci curvature of graphs . Tohoku Mathematical Journal, 63 0 (4): 0 605 -- 627, 2011

  19. [27]

    H., Jia, Z., and Yu, P

    Liu, K., Dou, Y., Zhao, Y., Ding, X., Hu, X., Zhang, R., Ding, K., Chen, C., Peng, H., Shu, K., Sun, L., Li, J., Chen, G. H., Jia, Z., and Yu, P. S. Bond: Benchmarking unsupervised outlier node detection on static attributed graphs. In Advances in Neural Information Processing...

  20. [28]

    Sphereface: Deep hypersphere embedding for face recognition

    Liu, W., Wen, Y., Yu, Z., Li, M., Raj, B., and Song, L. Sphereface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 212--220, 2017

  21. [29]

    Pick and choose: A gnn-based imbalanced learning approach for fraud detection

    Liu, Y., Ao, X., Qin, Z., Chi, J., Feng, J., Yang, H., and He, Q. Pick and choose: A gnn-based imbalanced learning approach for fraud detection. In Proceedings of the Web Conference 2021, 2021

  22. [30]

    McAuley, J. J. and Leskovec, J. From amateurs to connoisseurs: modeling the evolution of user expertise through online reviews. In WWW, 2013

  23. [31]

    Network alignment by discrete ollivier-ricci flow

    Ni, C.-C., Lin, Y.-Y., Gao, J., and Gu, X. Network alignment by discrete ollivier-ricci flow. In International symposium on graph drawing and network visualization, pp.\ 447--462. Springer, 2018

  24. [32]

    Ricci curvature of metric spaces

    Ollivier, Y. Ricci curvature of metric spaces. Comptes Rendus Mathematique, 345 0 (11): 0 643--646, 2007

  25. [33]

    Ricci curvature of markov chains on metric spaces

    Ollivier, Y. Ricci curvature of markov chains on metric spaces. Journal of Functional Analysis, 256 0 (3): 0 810--864, 2009

  26. [34]

    and Rossi, F

    Piccoli, B. and Rossi, F. On properties of the generalized wasserstein distance. Archive for Rational Mechanics and Analysis, 222: 0 1339--1365, 2016

  27. [35]

    A critical look at the evaluation of gnns under heterophily: are we really making progress? In ICLR, 2023

    Platonov, O., Kuznedelev, D., Diskin, M., Babenko, A., and Prokhorenkova, L. A critical look at the evaluation of gnns under heterophily: are we really making progress? In ICLR, 2023

  28. [36]

    and Akoglu, L

    Rayana, S. and Akoglu, L. Collective opinion spam detection: Bridging review networks and metadata. In KDD, pp.\ 985--994, 2015

  29. [37]

    Gad-nr : Graph anomaly detection via neighborhood reconstruction

    Roy, A., Shu, J., Li, J., Yang, C., Elshocht, O., Smeets, J., and Li, P. Gad-nr : Graph anomaly detection via neighborhood reconstruction. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, 2024

  30. [38]

    Karate club: an api oriented open-source python framework for unsupervised learning on graphs

    Rozemberczki, B., Kiss, O., and Sarkar, R. Karate club: an api oriented open-source python framework for unsupervised learning on graphs. In Proceedings of the 29th ACM international conference on information & knowledge management, pp.\ 3125--3132, 2020

  31. [39]

    Representation tradeoffs for hyperbolic embeddings

    Sala, F., De Sa, C., Gu, A., and R \'e , C. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, pp.\ 4460--4469. PMLR, 2018

  32. [40]

    Singh, K. V. and Vig, L. Improved prediction of missing protein interactome links via anomaly detection. Applied Network Science, 2: 0 1--20, 2017

  33. [41]

    and Knopp, P

    Sinkhorn, R. and Knopp, P. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21 0 (2): 0 343--348, 1967

  34. [42]

    Rethinking graph neural networks for anomaly detection

    Tang, J., Li, J., Gao, Z., and Li, J. Rethinking graph neural networks for anomaly detection. In International Conference on Machine Learning, 2022

  35. [43]

    Ricci curvatures of contact riemannian manifolds

    Tanno, S. Ricci curvatures of contact riemannian manifolds. Tohoku Mathematical Journal, Second Series, 40 0 (3): 0 441--448, 1988

  36. [44]

    Curvature-based clustering on graphs

    Tian, Y., Lubberts, Z., and Weber, M. Curvature-based clustering on graphs. arXiv preprint arXiv:2307.10155, 2023

  37. [45]

    P., Dong, X., and Bronstein, M

    Topping, J., Di Giovanni, F., Chamberlain, B. P., Dong, X., and Bronstein, M. M. Understanding over-squashing and bottlenecks on graphs via curvature. arXiv preprint arXiv:2111.14522, 2021

  38. [46]

    Ungar, A. A. Hyperbolic trigonometry and its application in the poincar \'e ball model of hyperbolic geometry. Computers & Mathematics with Applications, 41 0 (1-2): 0 135--147, 2001

  39. [47]

    Geometry of laplace-beltrami operator on a complete riemannian manifold

    Urakawa, H. Geometry of laplace-beltrami operator on a complete riemannian manifold. Progress in differential geometry, 22: 0 347--406, 1993

  40. [48]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph attention networks. arXiv:1710.10903, 2017

  41. [49]

    and Zhu, H

    Wang, C. and Zhu, H. Wrongdoing monitor: A graph-based behavioral anomaly detection in cyber security. IEEE Transactions on Information Forensics and Security, 17: 0 2703--2718, 2022

  42. [50]

    Decoupling representation learning and classification for gnn-based anomaly detection

    Wang, Y., Zhang, J., Guo, S., Yin, H., Li, C., and Chen, H. Decoupling representation learning and classification for gnn-based anomaly detection. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\ 1239--1248, 2021

  43. [51]

    Weber, M., Domeniconi, G., Chen, J., Weidele, D. K. I., Bellei, C., Robinson, T., and Leiserson, C. E. Anti-money laundering in bitcoin: Experimenting with graph convolutional networks for financial forensics. arXiv preprint arXiv:1908.02591, 2019

  44. [52]

    Pseudo-riemannian graph convolutional networks

    Xiong, B., Zhu, S., Potyka, N., Pan, S., Zhou, C., and Staab, S. Pseudo-riemannian graph convolutional networks. Advances in Neural Information Processing Systems, 35: 0 3488--3501, 2022

  45. [53]

    Casflow: Exploring hierarchical structures and propagation uncertainty for cascade prediction

    Xu, X., Zhou, F., Zhang, K., Liu, S., and Trajcevski, G. Casflow: Exploring hierarchical structures and propagation uncertainty for cascade prediction. IEEE Transactions on Knowledge and Data Engineering, 35 0 (4): 0 3484--3499, 2021

  46. [54]

    Hyperbolic graph attention network

    Zhang, Y., Wang, X., Shi, C., Jiang, X., and Ye, Y. Hyperbolic graph attention network. IEEE Transactions on Big Data, 8 0 (6): 0 1690--1701, 2021

  47. [55]

    Error-bounded graph anomaly loss for gnns

    Zhao, T., Deng, C., Yu, K., Jiang, T., Wang, D., and Jiang, M. Error-bounded graph anomaly loss for gnns. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pp.\ 1873--1882, 2020

  48. [56]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.