Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Partitioning Message Passing for Graph Fraud Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A new scheme that partitions neighbors by class and gives each node its own spectral filter substantially improves graph fraud detection.

desk verdict Useful spatial message-passing design for fraud detection with strong empirical results, but the spectral-filter theorem in Section 4 is not proven and should be fixed or dropped. read the letter →

arxiv 2412.00020 v1 pith:DMCSAJLD submitted 2024-11-16 cs.AI cs.LGcs.SI

Wei Zhuo , Zemin Liu , Bryan Hooi , Bingsheng He , Guang Tan , Rizal Fathony , Jia Chen This is my paper · ORCID
classification cs.AIcs.LGcs.SI
keywords graphfrauddetectionmessagepassingheterophilylabelimbalancespectralfilternode-specificneuralnetworks
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 proposes Partitioning Message Passing (PMP), a GNN aggregation scheme for graph fraud detection that does not try to remove heterophilic neighbors, as prior methods do, but instead aggregates fraud, benign, and unlabeled neighbors with separate node-specific transformations. The authors argue that the core problem is not heterophily itself but label imbalance: benign neighbors dominate the gradient because they share one transformation with fraud neighbors. PMP lets each center node adaptively weight the information from each neighbor class, including a learned mix for unlabeled neighbors, and the paper proves, under Theorem 1, that this spatial scheme is equivalent to an adaptive node-specific spectral graph filter. On four public fraud datasets and one industrial graph, PMP reports higher AUC, F1-Macro, and G-Mean than eleven published baselines.

What carries the argument

The load-bearing mechanism is the partitioning of the neighbor aggregation step by class labels, formalized with diagonal mask matrices $F$ and $B$ for fraud and benign training nodes. For a center node $v_i$, the unlabeled-neighbor mask is $(I - F - B)$, and its transformation uses a convex combination $\alpha_i W_{fr} + (1-\alpha_i)W_{be}$ where $\alpha_i$ is produced by a shared MLP from the node's own features. Node-specific weight generators $\Psi_{fr}(h_i)$ and $\Psi_{be}(h_i)$ output the class-specific transformations, so each node effectively receives its own filter. Theorem 1 then rewrites the PMP update as $(I-L)K(v_i)XW_{fr} + (I-L)(I-K(v_i))XW_{be}$ and claims that this equals $U g^i_{fr}(\Lambda)U^\top X W_{fr} + U g^i_{be}(\Lambda)U^\top X W_{be}$, defining the node-specific spectral filters.

What would settle it

Train a single-layer PMP on a small graph where the label mask $K(v_i)$ is known not to commute with the normalized Laplacian, compute the exact output of the spatial update, and compare it with the spectral formula $U g^i_{fr}(\Lambda)U^\top X W_{fr} + U g^i_{be}(\Lambda)U^\top X W_{be}$ from Theorem 1. If the two matrices differ on any entry, the claimed equivalence fails; equivalently, measure the frequency response of the trained layer on each node and check whether it matches the piecewise-defined $g^i(\Lambda)$.

Watch

Extended reading notes

Core claim

The central claim is that distinguishing neighbors by class during message passing—rather than excluding heterophilic neighbors or reweighting edges—is the key to applying GNNs to fraud detection. Concretely, PMP partitions each node's neighborhood into fraud, benign, and unlabeled sets, applies aggregation functions with distinct weight matrices to each set, and generates those weight matrices per center node from its own features. The paper's theoretical contribution is the claim that this scheme acts as a node-specific spectral graph filter: each node gets its own frequency response $g^i_{fr}(\Lambda)$ and $g^i_{be}(\Lambda)$, so the model can act as a low-pass filter in homophilic regions and a high-pass or band-pass filter in heterophilic regions. As a corollary, the paper argues that a single shared graph filter across all nodes cannot balance homophily and heterophily, and that PMP's node-level adaptivity is what yields its empirical gains.

Load-bearing premise

The spectral-filter theorem assumes the per-node label mask can be interchanged with the graph's diffusion operator, which only holds when the mask and Laplacian share eigenvectors; that commutation generally fails, so the derived filter formula may not describe the actual PMP computation.

Editorial extensions

If this is right

  • Fraud detection GNNs can keep the original graph and still handle heterophily, since class-aware aggregation replaces edge pruning or reweighting.
  • PMP is trainable with mini-batches, so it scales to graphs with tens of millions of edges where spectral baselines run out of memory or time.
  • The node-specific filter interpretation gives a principled explanation for why mixing homophily and heterophily helps: each node can choose its own frequency response.
  • The ablation shows that the simple partition step yields most of the gain, suggesting that class-distinct transformations are more important than elaborate sampling or reweighting modules.

Reading between the lines

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

  • The same partitioning idea could be applied to other imbalanced node-classification tasks beyond fraud, such as anomaly detection or rare-disease prediction, where the minority class is also heterophilically embedded.
  • The claim that each node has its own spectral filter is stronger than what the proof supports; a direct empirical check of the filter responses on real graphs would clarify whether the node specificity is real or an artifact of the commuting assumption.
  • Because PMP relies on labeled neighbors to define partitions, its benefit may shrink as label sparsity increases; extending the adaptive combination to purely unsupervised settings would test the limits of the approach.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Partitioning Message Passing (PMP), a GNN layer that aggregates fraud, benign, and unlabeled neighbors with distinct, node-adaptive weight matrices. It claims a theoretical result (Theorem 1) that PMP is equivalent to a node-specific spectral graph filter, and reports experiments on four public benchmarks plus an industrial graph showing gains over eleven baselines. The empirical evaluation reports AUC, F1-Macro, and G-Mean with mean/std over 10 trials and includes ablations.

Significance. If the theoretical claim were correct, it would provide a principled explanation for why label-partitioned message passing handles mixed homophily-heterophily and class imbalance. The empirical contribution is solid: PMP outperforms strong baselines on Yelp, Amazon, T-Finance, and T-Social under the supervised setting, with clear ablations and open-source code. However, the central theoretical assertion in Section 4 and the abstract is not supported by the proof in Appendix B, so the paper currently overstates its contribution.

major comments (3)
  1. [Appendix B, Eq. (8)] The expansion of the Wbe term is algebraically incorrect. Starting from FXWfr + BXWbe + (I-F-B)X(αiWfr + (1-αi)Wbe), the coefficient of Wbe should be (1-αi)I - (1-αi)F + αiB, but Eq. (8) writes B + (1-αi)I - (1-αi)B - (1-αi)B, which equals (1-αi)I + (2αi-1)B and differs from the correct expression by terms involving F. Consequently the subsequent identification of K(vi) and I-K(vi) in Eq. (9) is not a valid reformulation of the PMP update.
  2. [Appendix B, Eq. (10)] The step (I-L)K(vi) = U gi_fr(Λ) U^T presupposes that K(vi) is diagonalized by the Laplacian eigenvectors U, i.e., that K(vi) commutes with L. The proof does not establish this, and for a generic label mask K(vi) with 0/1/α entries this commutativity fails. Without it, the claimed spectral filter form in Eq. (6)-(7) does not follow.
  3. [Section 4, Eq. (7)] The filter coefficients are defined by conditions such as 'vj ∈ Nfr(vi)', which refer to the node index j, while Λ is indexed by the eigenvalues/eigenvectors of the Laplacian. There is no fixed correspondence between node order and eigenvector order, so these expressions do not define a well-defined diagonal function of Λ. This is a second, independent reason the theorem's statement is not established.
minor comments (4)
  1. [Section 5.2, Tables 6-7] The statement that PMP 'consistently surpasses baseline performances across almost all datasets and metrics' is too strong for the semi-supervised setting: on T-Finance (Table 7), H2-FDetector achieves AUC 94.37 versus PMP's 93.78, and on Amazon (Table 6) GHRN has F1-Macro 89.16 versus PMP's 87.72.
  2. [Table 3] The '+ + +root-specific weights' step decreases AUC on Amazon relative to '+ +adaptive combination' (97.57 vs 97.61), so the claim that the final component 'enhances performance on most metrics in most datasets' should be stated as 'most' or quantified.
  3. [Appendix B] The proof defines K(vi) as node-specific but then treats it as a single matrix in the spectral derivation; the notation should clarify that the matrix is applied for each center node, or the argument should be rewritten in terms of a mask operator to avoid index ambiguity.
  4. [Appendix A] The stated time complexity transitions from O(LN d + L|E|d^2) to O(LN d + (L|E|+N)d^2) without explanation; the latter appears to include the weight generators, but the derivation is implicit.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1's adaptive spectral filter is defined from the spatial label partition, so the spectral characterization is a reformulation by construction rather than an independent derivation.

  1. self definitional [Section 4 (Theorem 1, Eqs. 6-7) and Appendix B (Eq. 10)]
    "the node-specific convolution filters gi_fr(Λ) and gi_be(Λ) can be derived as Eq. (7). ... where the spectral convolution filters are diagonal matrices defined as: gi_fr(Λ)[j, j] = 1 − λj vj ∈ Nfr(vi); 0 vj ∈ Nbe(vi); αi(1 − λj) otherwise ... H(vi) = (I − L)K(vi)XWfr + (I − L)(I − K(vi))XWbe = Ugi_fr(Λ)U⊤XWfr + Ugi_be(Λ)U⊤XWbe."

    The filter entries in Eq. (7) are defined directly by the same neighbor-label partition Nfr(vi)/Nbe(vi) and the same node-specific scalar α_i that already constitute the spatial PMP update in Eqs. (3)-(5). Eq. (10) then asserts that this spatially defined update equals U g(Λ)U^T applied to the feature matrices, but g was chosen precisely to encode the diagonal label mask K(vi). Thus the spectral form is a notational restatement of the spatial model, not a consequence derived from the Laplacian's eigenstructure. Moreover, the equality in Eq. (10) requires (I−L)K(vi) to be simultaneously diagonalizable with L, i.e., K(vi) to commute with L, which generically fails; without that commutation, the diagonal-filter claim is even more clearly a stipulation.

full rationale

The empirical program of the paper is self-contained: PMP is evaluated against 11 baselines on held-out test splits across multiple datasets with repeated seeds, and the ablation study decomposes the contribution of each design choice. None of these results is a fitted quantity relabeled as a prediction. There is also no load-bearing self-citation: prior work by the authors appears only as related-work examples, and no uniqueness theorem is imported from their own papers. The single circular element is the theoretical claim in Theorem 1. In Eq. (7), the 'node-specific spectral filters' are defined entry-wise by whether a node v_j lies in Nfr(vi) or Nbe(vi) and by the same node-specific α_i that already defines the spatial aggregation in Eqs. (3)-(5). Eq. (10) then asserts the spatial update equals U g(Λ)U^T applied to XW; since g is chosen to encode the label mask K(vi), the spectral form is a notational restatement of the spatial model, not a result derived from the Laplacian's eigenstructure. Additionally, the equality in Eq. (10) requires K(vi) to commute with L, which generally fails, and even if that gap were repaired, the filter definition already contains the theorem's conclusion. Because the headline theoretical result is thus true by construction while the empirical claims remain independent, the paper exhibits partial circularity rather than a fully circular derivation.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The model introduces no new physical entities; it introduces learnable components (weight generators and the alpha blending function). The main formal burden is the spectral theorem, which relies on an unjustified commutativity assumption.

free parameters (6)
  • learning_rate = 0.01 for Yelp, Amazon, T-Finance; 0.001 for T-Social
    Tuned per dataset and reported in Table 5; affects optimization and final performance.
  • weight_decay = 0 for all datasets in Table 5
    Tuned per dataset; standard regularization hyperparameter.
  • dropout = 0 for Yelp, T-Finance, T-Social; 0.6 for Amazon
    Tuned per dataset; affects generalization.
  • hidden_dimension = 256 Yelp, 128 Amazon, 64 T-Finance, 128 T-Social
    Tuned per dataset; controls model capacity.
  • batch_size = 512 Yelp, 128 Amazon, 256 T-Finance, 217 T-Social
    Tuned per dataset; affects training dynamics and scalability.
  • number_of_layers = 1 for all datasets
    Chosen via parameter study in Appendix D.5; L=1 is optimal.
assumptions (3)
  • domain assumption The graph is undirected so that the symmetric normalized Laplacian has an eigendecomposition L = U Λ U^T.
    Theorem 1 states the graph is undirected. Fraud graphs are often directed (e.g., transaction networks), so this limits the scope of the spectral claim.
  • ad hoc to paper The label mask K(vi) is diagonal in the eigenbasis of the Laplacian, allowing (I-L)K(vi) to be expressed as U g(Λ) U^T with g depending on node indices.
    This assumption is used implicitly in the proof of Theorem 1 (Appendix B, Eq. 10) but is false in general. It is needed for the spectral filter derivation to go through.
  • domain assumption For the influence analysis, the graph is regular so all non-zero off-diagonal entries of Ahat^k are equal.
    Stated in Section 2.2 as a simplifying assumption to isolate the label imbalance effect. Real fraud graphs are not regular.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Partitioning Message Passing for Graph Fraud Detection." pith.science (2026). https://pith.science/paper/DMCSAJLD

@misc{pith2026241200020,
  author       = {Pith},
  title        = {Pith review of: Partitioning Message Passing for Graph Fraud Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMCSAJLD}},
  note         = {Machine review of arXiv:2412.00020}
}
read the original abstract

Label imbalance and homophily-heterophily mixture are the fundamental problems encountered when applying Graph Neural Networks (GNNs) to Graph Fraud Detection (GFD) tasks. Existing GNN-based GFD models are designed to augment graph structure to accommodate the inductive bias of GNNs towards homophily, by excluding heterophilic neighbors during message passing. In our work, we argue that the key to applying GNNs for GFD is not to exclude but to {\em distinguish} neighbors with different labels. Grounded in this perspective, we introduce Partitioning Message Passing (PMP), an intuitive yet effective message passing paradigm expressly crafted for GFD. Specifically, in the neighbor aggregation stage of PMP, neighbors with different classes are aggregated with distinct node-specific aggregation functions. By this means, the center node can adaptively adjust the information aggregated from its heterophilic and homophilic neighbors, thus avoiding the model gradient being dominated by benign nodes which occupy the majority of the population. We theoretically establish a connection between the spatial formulation of PMP and spectral analysis to characterize that PMP operates an adaptive node-specific spectral graph filter, which demonstrates the capability of PMP to handle heterophily-homophily mixed graphs. Extensive experimental results show that PMP can significantly boost the performance on GFD tasks.

Figures

Figures reproduced from arXiv: 2412.00020 by the authors.

Figure 1
Figure 1. Comparison between generic message passing GNN and partitioning message passing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) A Barabasi–Albert graph with 500 nodes, with 10% nodes are fraud. Features of benign ´ nodes (depicted in blue) follow a Gaussian distribution, N (1, 1), while those of fraud nodes (shown in red) are drawn from N (5, 1). (b) Spectral convolution filters of PMP for a random node vi . and W(l) be,i tailored to its unique features, while the generation process MLPfr and MLPbe remains consistent and is governed by s… view at source ↗
Figure 4
Figure 4. Influence distribution. and standard deviation for each model as benchmarks for comparison. We provide the detailed hyperparameter tuning strategies of baselines and hyperparameter setting of PMP in Appendix D.4. 5.2 PERFORMANCE COMPARISON 2000 4000 6000 8000 10000 Testing Time (s) 70 75 80 85 90 95 AUC CARE-GNN PC-GNN BWGNN GraphSAGE GHRN GDN PMP [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: AUC vs. testing time on T￾Social. For public datasets, results derived from the supervised setting can be found in Tables 1 and 2, while those from the semi-supervised setting are detailed in Tables 6 and 7 of Appendix E.1. For the Grab dataset, results are shown in Ap…
Figure 5
Figure 5. Figure 5: Label distributions of the labeled neighborhoods of the training nodes. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Sensitivity study of different hyperparameter combinations. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Influence distribution. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs

    cs.LG 2025-07 conditional novelty 5.0 of 10

    DGP condenses neighbors into coarse-grained summaries while preserving the target node's full text, improving LLM-based fraud detection by up to 6.8 absolute AUPRC points over state-of-the-art methods.

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [5]

    Intention-aware heterogeneous graph attention networks for fraud transactions detection

    Can Liu, Li Sun, Xiang Ao, Jinghua Feng, Qing He, and Hao Yang. Intention-aware heterogeneous graph attention networks for fraud transactions detection. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp. 3280–3288, 2021a. Meng Liu, Zhengyang Wang, and Shuiwang Ji. Non-local graph neural networks. IEEE transactions on...

    arXiv 2021

  2. [6]

    Is heterophily a real nightmare for graph neural networks to do node classification? arXiv preprint arXiv:2109.05641,

    Sitao Luan, Chenqing Hua, Qincheng Lu, Jiaqi Zhu, Mingde Zhao, Shuyuan Zhang, Xiao-Wen Chang, and Doina Precup. Is heterophily a real nightmare for graph neural networks to do node classification? arXiv preprint arXiv:2109.05641,

    arXiv

  3. [12]

    Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou, Shuang Yang, and Yuan Qi

    URL https://openreview.net/forum?id=rJXMpikCZ. Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou, Shuang Yang, and Yuan Qi. A semi-supervised graph attentive network for financial fraud detection. In 2019 IEEE International Conference on Data Mining (ICDM), pp. 598–607. IEEE,

    work page 2019

  4. [13]

    Label information enhanced fraud detection against low homophily in graphs

    Yuchen Wang, Jinghui Zhang, Zhengjie Huang, Weibin Li, Shikun Feng, Ziheng Ma, Yu Sun, Dianhai Yu, Fang Dong, Jiahui Jin, et al. Label information enhanced fraud detection against low homophily in graphs. In Proceedings of the ACM Web Conference 2023, pp. 406–416,

    work page 2023

  5. [15]

    Interpreting and unifying graph neural networks with an optimization framework

    Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. Interpreting and unifying graph neural networks with an optimization framework. In Proceedings of the Web Conference 2021 , pp. 1215–1226,

    work page 2021

  6. [16]

    Diffusiongan: Network embedding for informa- tion diffusion prediction with generative adversarial nets

    Wei Zhuo, Yanan Zhao, Qianyi Zhan, and Yuan Liu. Diffusiongan: Network embedding for informa- tion diffusion prediction with generative adversarial nets. In 2019 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computing & Communications, Social Computing & Networking (ISPA/BDCloud/SocialCom/Su...

    work page 2019

  7. [18]

    For the sake of simplicity, we analyze the first layer of PMP as an example and omit superscripts

    Then the mask matrix of unlabeled nodes is thus I − F − B. For the sake of simplicity, we analyze the first layer of PMP as an example and omit superscripts. Specifically, the feature transformation step of Eq. (3) can be reformulated as: FXWfr + BXWbe + (I − F − B)X (αiWfr + (1 − αi)Wbe) = (F + αiI − αiF − αiB) XWfr + (B + (1 − αi)I − (1 − αi)B − (1 − αi...

    work page 2017

  8. [19]

    C, representing the number of classes, is set to 2 for GFD

    is defined as: ˆη = 1 C − 1 C−1X k=0 ηk − |Ck| N + (11) where [·]+ = max(·, 0). C, representing the number of classes, is set to 2 for GFD. Ck denotes the set of nodes in class k, with k = 0 corresponding to benign nodes and k = 1 to fraud nodes. ηk is the 15 Published as a conference paper at ICLR 2024 class-wise homophily metric: ηk = P vi∈Ck |Nk(vi)|P ...

    work page 2024

Show all 22 references
  1. [20]

    Due to the anonymity and privacy policy, we exclude the exact details of the graph but roughly the graph includes over 1 million nodes and over 10 million edges

    The industrial graph includes the transaction of a month in a leading super app. Due to the anonymity and privacy policy, we exclude the exact details of the graph but roughly the graph includes over 1 million nodes and over 10 million edges. Table 4: Summary of dataset statis...

    2024

  2. [21]

    Table 5: Hyperparameter settings for PMP. Dataset lr wd do bs L d ′ Yelp 0.01 0 0 512 1 256Amazon 0.01 0 0.6 128 1 128T-Finance 0.01 0 0.4 256 1 64 T-Social 0.001 0 0 217 1 128 D.5 P ARAMETER STUDY We investigate the sensitivity in relation to the key hyperparameters in our mo...

    2024

  3. [22]

    extends 18 Published as a conference paper at ICLR 2024 Table 8: Experiment Results on Grab (40% training ratio) Method Grab AUC F1-Macro G-Mean Care-GNN 99.58 ±0.01 68.87±0.21 98.31±0.03 PC-GNN 98.97 ±0.02 66.32±1.15 98.08±0.14 H2-FDetector OOT OOT OOT BWGNN 99.79 ±0.03 80.63...

    2024

  4. [2011]

    Geom-gcn: Geometric graph convolutional networks

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net,

    2020

  5. [2013]

    Cadue: Content- agnostic detection of unwanted emails for enterprise security

    11 Published as a conference paper at ICLR 2024 Mohamed Nabeel, Enes Altinisik, Haipei Sun, Issa Khalil, Hui Wang, and Ting Yu. Cadue: Content- agnostic detection of unwanted emails for enterprise security. In Proceedings of the 24th Interna- tional Symposium on Research in At...

    2024

  6. [2015]

    Modeling relational data with graph convolutional networks

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In The Semantic Web: 15th International Conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, Proceedings...

    2018

  7. [2016]

    Leyan Deng, Chenwang Wu, Defu Lian, Yongji Wu, and Enhong Chen

    URL https://arxiv.org/abs/1606.09375. Leyan Deng, Chenwang Wu, Defu Lian, Yongji Wu, and Enhong Chen. Markov-driven graph convolutional networksfor social spammer detection. IEEE Transactions on Knowledge and Data Engineering,

    arXiv

  8. [2017]

    New benchmarks for learning on non- homophilous graphs

    Derek Lim, Xiuyu Li, Felix Hohne, and Ser-Nam Lim. New benchmarks for learning on non- homophilous graphs. arXiv preprint arXiv:2104.01404,

    arXiv

  9. [2018]

    H2-fdetector: A gnn-based fraud detector with homophilic and heterophilic connections

    Fengzhao Shi, Yanan Cao, Yanmin Shang, Yuchen Zhou, Chuan Zhou, and Jia Wu. H2-fdetector: A gnn-based fraud detector with homophilic and heterophilic connections. In Proceedings of the ACM Web Conference 2022 (TheWebConf), pp. 1486–1494,

    2022

  10. [2019]

    13 Published as a conference paper at ICLR 2024 A A LGORITHMIC DETAILS Algorithm 1 PMP forward propatation Input: Fraud graph G = (V, Er, X, Y); Depth L; Batch size B; Output: Logits Z ∈ RN 1: for l ∈ {1, · · ·, L} do 2: for each batch ⊆ Gof size B do 3: for vi ∈ batch do 4: b...

    2024

  11. [2020]

    Ad- dressing heterophily in graph anomaly detection: A perspective of graph spectrum

    Yuan Gao, Xiang Wang, Xiangnan He, Zhenguang Liu, Huamin Feng, and Yongdong Zhang. Ad- dressing heterophily in graph anomaly detection: A perspective of graph spectrum. In Proceedings of the ACM Web Conference 2023, pp. 1528–1538, 2023a. Yuan Gao, Xiang Wang, Xiangnan He, Zhen...

    2023

  12. [2021]

    Hyunsoo Cho, Jinseok Seol, and Sang-goo Lee

    URL https://openreview.net/forum?id=n6jl7fLxrP. Hyunsoo Cho, Jinseok Seol, and Sang-goo Lee. Masked contrastive learning for anomaly detection. arXiv preprint arXiv:2105.08793,

    arXiv

  13. [2022]

    Shebuti Rayana and Leman Akoglu

    doi: 10.14778/3494124.3494128. Shebuti Rayana and Leman Akoglu. Collective opinion spam detection: Bridging review networks and metadata. In Proceedings of the 21th acm sigkdd international conference on knowledge discovery and data mining, pp. 985–994,

  14. [2023]

    A comprehensive survey on graph neural networks

    12 Published as a conference paper at ICLR 2024 Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 32(1):4–24,

    2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.