Pith. sign in

REVIEW 3 major objections 8 minor 90 references

Learnable Kernel Density Estimation for Graphs and Its Application to Graph-Level Anomaly Detection

T0 review · 3 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LGKDE learns a probability density over graphs by training a multi-scale kernel density estimator on a GNN-learned MMD metric, with density contrast against perturbed copies providing consistency guarantees and strong anomaly detection.

desk verdict Genuinely strong empirical anomaly detection and a sound engineering combination, but the consistency theorems rest on an unproven dint=1 assumption and a fixed-parameter analysis that never connects to the training objective. read the letter →

arxiv 2505.21285 v5 pith:CIN7UBEM submitted 2025-05-27 cs.LG stat.ML

Xudong Wang , Ziheng Sun , Chris Ding , Jicong Fan This is my paper
classification cs.LGstat.ML
keywords graphdensityestimationkernelmaximummeandiscrepancyneuralnetworksgraph-levelanomalydetectionmulti-scalekernelsconsistencyguarantees
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 attempts to establish that the probability density of a graph collection can be estimated by a deep, learnable kernel density estimator rather than by a fixed handcrafted graph kernel. The proposed LGKDE represents each graph as a distribution over GNN node embeddings, measures graph distances with maximum mean discrepancy, and builds a multi-scale kernel density estimate in that learned metric. All parameters, including the mixture weights over bandwidths, are trained by maximizing the density of each normal graph relative to its perturbed copies, where perturbations shuffle node features and modify the adjacency matrix's singular values. The authors prove consistency, a convergence rate, stability of the density estimate, and a generalization bound for this estimator, and they report that it recovers known synthetic distributions and outperforms existing graph anomaly detectors on most of twelve benchmark datasets. Establishing this matters because graph-level anomaly detection would then rest on an explicit density model with statistical guarantees instead of on ad-hoc embedding or hypersphere assumptions.

What carries the argument

The load-bearing machinery is a multi-scale kernel density estimator defined on a learned maximum mean discrepancy metric between graphs, together with a density-contrasting training objective. Each graph is encoded as a bag of GNN node embeddings; the MMD distance $d_{MMD}$ takes a supremum over a family of Gaussian kernels on those embeddings, giving a differentiable, permutation-invariant geometry. The KDE components use a Gaussian profile with normalization constant $C_{d_{\mathrm{int}}}=(2\pi)^{d_{\mathrm{int}}/2}$, where the paper sets $d_{\mathrm{int}}=1$ because each kernel acts on the scalar MMD distance. The objective $-\sum_i (\hat{f}_{KDE}(G_i)-\hat{f}_{KDE}(\tilde{G}_i^{(j)}))/\hat{f}_{KDE}(G_i)$, with perturbed copies $\tilde{G}$ generated by node-feature shuffling and energy-based spectral modification, drives the learned metric and bandwidth weights to assign high density only inside the normal region. The theoretical results—$L^1$ consistency, the MISE rate, Lipschitz stability of the density under metric and graph perturbations, and the generalization bound—all depend on this learned geometry.

What would settle it

Train LGKDE on synthetic graphs with a known density and estimate the intrinsic dimension of the learned MMD distance matrix (for example, by PCA on the distance Gram matrix or by fitting local covariance ranks). If that dimension is greater than one, the normalization constant $C_{d_{\mathrm{int}}}$ used in Eq. (8) is wrong and the claimed $L^1$ consistency and $O(N^{-0.8})$ MISE rate cannot hold; a complementary check is whether the learned density's peaks match the true mixture weights on a two-mode synthetic graph distribution.

Watch

Extended reading notes

Core claim

The central claim is that a proper density over graphs can be formed by a mixture of Gaussian KDE components evaluated on a learned MMD distance, with the density explicitly trained to place normal graphs above their perturbed counterparts. In the authors' framing, the estimator $\hat{f}_{KDE}(G)=\sum_{k=1}^{M}\pi_k(\alpha)\frac{1}{N}\sum_{i=1}^{N}K_{KDE}(d_{MMD}(G,G_i),h_k)$, with learnable softmax weights $\pi_k$, is statistically consistent when the bandwidths shrink appropriately, converges in mean integrated squared error at rate $O(N^{-4/(4+d_{\mathrm{int}})})$, and reaches $O(N^{-0.8})$ when the intrinsic dimension $d_{\mathrm{int}}=1$. The anomaly score is simply the negative density, with a threshold set by a percentile of reference-set densities. Empirically, LGKDE recovers the Beta(2,2) generative density of synthetic Erdős–Rényi graphs and achieves an average AUROC of 79.37% over twelve graph classification benchmarks.

Load-bearing premise

The proofs assume the learned graph metric behaves locally like a one-dimensional space, so the kernel normalizes with $\sqrt{2\pi}$ and the rate is $O(N^{-0.8})$, and they assume the optimal parameters are fixed; if the true intrinsic dimension is greater than one or the contrastive objective does not track the true density, the stated consistency and convergence guarantees fail.

Editorial extensions

If this is right

  • Graph-level anomaly detection becomes direct density thresholding: any query graph is scored by $\hat{f}_{KDE}$ and flagged anomalous when its score falls below the reference percentile $\tau$.
  • The $L^1$ consistency theorem implies that adding more normal graphs drives the learned density toward the true graph distribution, so the detector inherits standard statistical guarantees rather than relying on a hypersphere or cluster assumption.
  • The stability bounds imply that small perturbations of a normal graph change its density only slightly, which the authors use to argue for low false-positive rates and for tolerance to moderate training contamination (performance in the paper holds up to 10% contamination).
  • Because the metric and the KDE mixture weights are learned jointly, two-stage pipelines that first embed graphs and then fit a KDE are dominated by LGKDE; on MUTAG the paper reports 91.63% AUROC for LGKDE versus 81.94% for GAE followed by KDE.
  • The $O(N^{-0.8})$ rate means that, under the paper's one-dimensional assumption, the estimator is minimax-optimal in sample complexity and does not degrade with ambient node-feature dimension.

Reading between the lines

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

  • If the learned MMD metric's intrinsic dimension on a given dataset exceeds one, the $\sqrt{2\pi}$ normalization and the $O(N^{-0.8})$ rate are not justified; a practical user should estimate the local dimensionality of the learned distance matrix before relying on the stated guarantees.
  • The density-contrast objective treats perturbed graphs as soft negative evidence rather than as labeled anomalies, which suggests a natural extension to continually shifting distributions by resampling perturbations near the current decision boundary.
  • The learned metric itself is a reusable artifact: the same MMD distances could support out-of-distribution detection, clustering, or visualization on domains where annotated anomalies are unavailable.
  • The quadratic cost in the reference set is the main scaling bottleneck; the paper's own sampling experiments indicate that density-stratified or importance sampling of reference graphs could extend the method to much larger graph corpora with little performance loss.
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

3 major / 8 minor

Summary. The paper proposes LGKDE, an unsupervised learnable kernel density estimator for graphs. Graphs are embedded by a GNN into node-embedding distributions, pairwise distances are computed with the deep MMD metric of Sun and Fan (2024), and a multi-scale KDE with learnable softmax mixture weights is evaluated on those distances. Training maximizes, for each graph, the density of the original graph relative to the densities of its perturbed counterparts, where perturbations randomly permute node features and modify the adjacency spectrum via SVD-based edge addition/removal. The authors claim consistency (Theorem 4.1), a minimax-optimal MISE rate O(N^{-4/(4+dint)}) with O(N^{-0.8}) when dint = 1 (Theorem 4.2), Lipschitz robustness to metric and graph perturbations (Theorems 4.3-4.4, Corollary 4.5), and a generalization bound (Theorem 4.6). Empirically, the paper reports density-recovery experiments on synthetic ER/BA/WS/SBM graphs and anomaly detection on 12 TU benchmark datasets, where LGKDE obtains the best average AUROC (79.37%) and average rank (1.67) over five seeds, supported by ablations, a contamination-robustness analysis, and runtime comparisons.

Significance. If the theoretical claims were valid, the paper would make a useful contribution by connecting deep metric learning with classical KDE for a non-Euclidean domain and by providing formal guarantees for a graph-level density estimator. The empirical package is a genuine strength: twelve benchmark datasets, three metrics, five seeds, extensive ablations, a training-contamination analysis, a runtime study, and a falsifiable rate claim O(N^{-0.8}) are reported, and the best-average-AUROC result is credible. However, as explained in the major comments, the headline consistency and rate results are not established by the manuscript: the dimension entering the kernel normalization is asserted rather than derived or measured, and the proofs concern a fixed-parameter estimator rather than the contrastively trained one. The theoretical section therefore cannot be credited as support for the empirical results, and the contribution reduces to the algorithmic proposal plus its empirical evaluation, which is substantial but weaker than the paper's advertised guarantees.

major comments (3)
  1. [§3.3.3, Eq. (8); Appendices F.2-F.3] The central theoretical claim rests on the assertion that dint = 1 for the learned MMD metric, but the paper never establishes that the local volume-growth dimension of the base measure equals one. The normalization of a KDE kernel must match the local dimension D of the measure μ in the metric space: on a D-dimensional manifold, the integral of (1/(√(2π)h)) K0(d(G,G')²/h²) with respect to dμ(G') scales as O(h^{D-1}), because the μ-volume of a metric ball of radius h is O(h^D). Hence, if D > 1, the estimator's expectation tends to 0, so Theorem 4.1's L1 consistency fails and the claimed rate in Theorem 4.2 is not N^{-0.8}. The justification in Appendix F.3 ('the kernel mass concentrates along the single distance dimension') confuses the dimension of the kernel's argument with the dimension of the underlying measure; a radial kernel in R^D has a scalar argument but requires h^{-D} normalization. The proof of Theorem 4.1 in Appendix F.2 simultaneously uses ∫ K_h(G,G') dμ(G') → 1 and a Riemannian Taylor expansion with a Laplace-Beltrami operator, which is internally consistent only if the manifold dimension equals dint = 1. The t-SNE visualizations in Figure 5 and the synthetic experiments do not measure the metric dimension, so Theorems 4.1 and 4.2 are unsupported as stated.
  2. [§3.3.3, Eq. (9); Appendix F.2] The consistency and MISE proofs in Appendices F.2-F.3 analyze the estimator at fixed 'optimal' parameters (θ*, α*), but no result connects solutions of the contrastive objective in Eq. (9) to θ*, α*, or to the true density f*. The objective is self-referential: it trains the KDE to assign higher density to each training graph than to that graph's own perturbed copies, a property the model can satisfy by concentrating mass on the training set without approximating the data-generating measure P*. Therefore the guarantees proven for the fixed-parameter family do not apply to the estimator deployed in Algorithm 3, and the statement in §5.2.2 that the empirical AUROC results 'naturally align' with Theorems 4.1-4.2 is not supported by the manuscript. A theorem establishing, or at least characterizing, the identifiability and consistency of the density learned by Eq. (9) is needed for the paper's central claim.
  3. [Theorem 4.6, Eq. (15); Appendix F.5, Eq. (33)] The displayed generalization bound is dimensionally inconsistent as printed. The second term in Eq. (15) (and in Eq. (33) of the appendix), '8√(en)π c_h²', has no denominator, whereas the Rademacher derivation in Eqs. (25)-(27) of Appendix F.5 produces terms such as 4/N and quantities of the form (8 + 24abL_f √ln(2d²) ln N)/N. As printed, the term grows with the graph size n and the minimal bandwidth c_h, contradicting both the proof and the paper's claim that 'the graph node number n has little impact on the bound'. This is a concrete error in a theorem displayed in the main text; the second term presumably should be a constant divided by N, with the factor (√e n π c_h²)^{-1} carried by the Lipschitz constant L_f.
minor comments (8)
  1. [§3.3.3 and Appendix E.3] The anomaly threshold parameter is reported inconsistently: §3.3.3 and §5.2.1 state γTH = 0.1, while Appendix E.3 says 'we set γTH = 10'; please reconcile.
  2. [Theorem 4.3 proof, Eq. (20)] In Eq. (20) of Appendix F.4.1, the last line has a subscript typo: dMMD(G1, Gj) should be dMMD(G1, G2).
  3. [§3.3.1 and Algorithm 1] The perturbation scheme defines a fraction ppert of singular values to be modified, but Algorithm 1 neither takes ppert as an input nor uses it, and Eq. (4) modifies entire groups Sh or Sl; please clarify how ppert is intended to interact with the adaptive ratio r.
  4. [Appendix E.4.1] The break-even argument for the complexity comparison states that 'most benchmark graphs have more than 25 nodes', which is contradicted by Table 3 (MUTAG 17.93, AIDS 15.69, IMDB-B 19.77 average nodes); the comparison with SIGNET should be revisited with dataset-specific node counts.
  5. [§5.1 and Table 7] The synthetic density-recovery evidence is only qualitative: the average estimated densities in Table 7 are not normalized (they do not integrate to a constant over p), and no quantitative fit statistic such as KL divergence or correlation is reported; the qualitative peak near p = 0.5 is encouraging but is weaker evidence than the phrase 'successfully recovered the underlying distribution' suggests.
  6. [References] Several references need correction: the Beckman and Cook entry is corrupted ('Outlier. . . . . . . . . . s.Technometrics'), and the MMD paper by Gretton et al. is cited as a NIPS contribution although the canonical citation is JMLR 12 (2012).
  7. [Appendix E.3 and Listing 1] The code is withheld until acceptance ('Code will be public and open source after paper acceptance'), which is a limitation for a heavily empirical paper; releasing the full repository rather than only a snippet would substantially strengthen reproducibility.
  8. [Figure 2 and Appendix F.5] The caption of Figure 2 contains fragments of algorithm pseudocode and is garbled, and Appendix F.5 contains the typo 'it si 1-Lipschitz'; both should be fixed in revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical results are conditional standard KDE statements, the empirical evaluations are out-of-sample on fixed benchmarks, and the overlapping-author citation supplies independent support rather than a self-referential reduction.

full rationale

I walked the derivation chain and found no step in which a prediction or first-principles result reduces by construction to its inputs. The density estimator in Eqs. (7)-(8) is a standard multi-scale KDE defined on a learned MMD metric; Theorems 4.1 and 4.2 are proved in Appendix F under explicit bandwidth conditions and Assumption F.1 (bounded GNN weights and smooth true density), not under fitted values or under the contrastive objective itself. The rate O(N^{-0.8}) follows from the stated dint = 1 specialization; the claim that dint = 1 is an unsupported geometric assertion in Appendix F.3, but that is an assumption whose failure would invalidate the theorem, not a circular reduction of the theorem to its own conclusion. The training objective Eq. (9) maximizes the density of normal graphs relative to self-generated perturbations, but the anomaly detection results are measured on held-out minority-class graphs from standard benchmarks, so the reported AUROC is not statistically forced by the fit. The synthetic density experiments similarly check that the learned density tracks the known generating parameter; they are not a fitted quantity renamed as a prediction. The only load-bearing citation to overlapping-author work is Proposition F.5 from Sun and Fan (2024), used in the proof of Theorem 4.4; it is a parameter-free perturbation bound with stated assumptions that do not include the present paper's target results, and it is not invoked as a uniqueness theorem to exclude alternatives. Accordingly, the paper is self-contained against external benchmarks and no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. The central derivation depends on four classes of assumptions: standard kernel/MMD theory, smoothness of a density on graph space, the ad hoc dint=1 intrinsic-dimension choice, and the belief that perturbations have lower true density. The learned parameters and hand-set hyperparameters listed above are fitted or chosen by hand and are load-bearing for the empirical results.

free parameters (4)
  • GNN parameters theta = learned on training data
    Define the MMD metric; the consistency theory assumes a fixed optimal theta* without proving convergence of the training objective.
  • KDE mixture weights alpha = learned via softmax over M=5 bandwidths
    Adaptively weight the multi-scale density components; fitted to data by Eq. (9).
  • Bandwidth set HKDE = 10^-2, 10^-1, 10^0, 10^1, 10^2
    Hand-chosen logarithmic grid; ablations show strong sensitivity to bandwidth choice.
  • Perturbation hyperparameters = rswap, ppert, tau1=0.5, tau2=0.75, rmax=10, gamma_TH=0.1
    Hand-chosen; no sensitivity analysis reported for rswap or ppert, yet they define the contrastive training signal.
assumptions (4)
  • domain assumption Bounded GNN weights and smoothness of the true density f* (Assumption F.1)
    Invoked in all proofs; the smoothness of a density on the abstract graph space is assumed, not established.
  • ad hoc to paper The scalar MMD distance induces a one-dimensional intrinsic manifold (dint=1)
    Controls KDE normalization and yields the claimed O(N^{-0.8}) rate; no derivation is provided.
  • domain assumption Perturbed versions of normal graphs have lower density under the true distribution
    This is the rationale for the contrastive objective in Eq. (9); if false, the training signal is misaligned with density.
  • standard math The MMD kernel family Kemb is characteristic and yields a valid metric
    Standard kernel two-sample test theory; needed for the metric property and the triangle inequality in the robustness proofs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learnable Kernel Density Estimation for Graphs and Its Application to Graph-Level Anomaly Detection." pith.science (2026). https://pith.science/paper/CIN7UBEM

@misc{pith2026250521285,
  author       = {Pith},
  title        = {Pith review of: Learnable Kernel Density Estimation for Graphs and Its Application to Graph-Level Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIN7UBEM}},
  note         = {Machine review of arXiv:2505.21285}
}
read the original abstract

This work proposes a framework LGKDE that learns kernel density estimation for graphs. The key challenge in graph density estimation lies in effectively capturing both structural patterns and semantic variations while maintaining theoretical guarantees. Combining graph kernels and kernel density estimation (KDE) is a standard approach to graph density estimation, but has unsatisfactory performance due to the handcrafted and fixed features of kernels. Our method LGKDE leverages graph neural networks to represent each graph as a discrete distribution and utilizes maximum mean discrepancy to learn the graph metric for multi-scale KDE, where all parameters are learned by maximizing the density of graphs relative to the density of their well-designed perturbed counterparts. The perturbations are conducted on both node features and graph spectra, which helps better characterize the boundary of normal density regions. Theoretically, we establish consistency and convergence guarantees for LGKDE, including bounds on the mean integrated squared error, robustness, and generalization. We validate LGKDE by demonstrating its effectiveness in recovering the underlying density of synthetic graph distributions and applying it to graph anomaly detection across diverse benchmark datasets. Extensive empirical evaluation shows that LGKDE demonstrates superior performance compared to state-of-the-art baselines on most benchmark datasets.

Figures

Figures reproduced from arXiv: 2505.21285 by the authors.

Figure 1
Figure 1. shows t-SNE [Van der Maaten and Hinton, 2008] visualizations of kernel matrices computed using different methods on the MUTAG dataset. Traditional graph kernels like WL and PK struggle to effectively separate graphs from different classes while our LGKDE learned kernel achieves clear separation between classes while maintaining smooth transitions in the metric space. This comparison underscores the importance of lea… view at source ↗
Figure 2
Figure 2. Framework of our proposed LGKDE These challenges motivate our development of LGKDE, a scalable kernel density estimation framework that effectively models the distributions of graphs through deep learning. 3.2 Overview of LGKDE Framework Since graphs are non-Euclidean data and often very complex, it is non-trivial to train a deep learning model to estimate the density of graphs when there is no available supervision… view at source ↗
Figure 3
Figure 3. Demonstration of the node feature and energy￾based spectral perturbation on a MUTAG molecule. Left: original graph; Middle: edge removal through high￾energy group division; Right: edge addition through low-energy group multiplication. See more case studies in Appendix D. Inspired by [Sun and Fan, 2024], we use a deep graph MMD model to compute meaning￾ful distances between graphs. The key idea is to represent each g… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Density estimation results on synthetic Erdos–Rényi graphs. Left: Ground truth Beta(2, 2) distribution ˝ for edge probability p. Middle: Learned density estimate from LGKDE versus edge probability p, showing the expected peak around p = 0.5. Right: Distribution of node…
Figure 5
Figure 5. Figure 5: t-SNE visualization (perplexity=30) of learned kernel matrix on the MUTAG dataset. For [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Ablation Studies on MUTAG Dataset B.5.1 Validation on Synthetic Erdos–Rényi Graphs ˝ To directly assess whether LGKDE learns the underlying probability density function (PDF) of graphs, we performed experiments on synthetically generated Erdos–Rényi (ER) graphs with kn…
Figure 7
Figure 7. Figure 7: Density estimation results on synthetic Erd [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Density gap analysis on MUTAG. Shows the average density assigned to true normal [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Visualization of MUTAG graphs. (a) The graph with minimum average MMD distance to [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Spectral Perturbation: Add Edges Mode via Amplify [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Spectral Perturbation: Remove Edges Mode via Shrink [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: shows that the histogram of edge change ratio of spectral perturbation on the whole [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

90 extracted references · 60 canonical work pages

  1. [1]

    Graph based anomaly detection and description: a survey

    Leman Akoglu, Hanghang Tong, and Danai Koutra. Graph based anomaly detection and description: a survey. Data mining and knowledge discovery , 29(3):626--688, 2015

    2015

  2. [2]

    Enhancing one-class support vector machines for unsupervised anomaly detection

    Mennatallah Amer, Markus Goldstein, and Slim Abdennadher. Enhancing one-class support vector machines for unsupervised anomaly detection. In Proceedings of the ACM SIGKDD workshop on outlier detection and description , pages 8--15, 2013

    2013

  3. [3]

    Theoretical numerical analysis , volume 39

    Kendall Atkinson and Weimin Han. Theoretical numerical analysis , volume 39. Springer, 2005

    2005

  4. [4]

    Emergence of scaling in random networks

    Albert-L \'a szl \'o Barab \'a si and R \'e ka Albert. Emergence of scaling in random networks. science , 286(5439):509--512, 1999

    1999

  5. [5]

    Outliers in statistical data , volume 3

    Vic Barnett, Toby Lewis, et al. Outliers in statistical data , volume 3. Wiley New York, 1994

    1994

  6. [6]

    Spectrally-normalized margin bounds for neural networks

    Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. Advances in neural information processing systems , 30, 2017

    2017

  7. [7]

    Outlier………

    Richard J Beckman and R Dennis Cook. Outlier………. s. Technometrics , 25(2):119--149, 1983

    1983

  8. [8]

    Shortest-path kernels on graphs

    Karsten M Borgwardt and Hans-Peter Kriegel. Shortest-path kernels on graphs. In Fifth IEEE international conference on data mining , pages 74--81, 2005

    2005

Show all 90 references
  1. [9]

    Lof: identifying density-based local outliers

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and J \"o rg Sander. Lof: identifying density-based local outliers. In Proceedings of the 2000 ACM SIGMOD international conference on Management of data , pages 93--104, 2000

    2000

  2. [10]

    Lg-fgad: An effective federated graph anomaly detection framework

    Jinyu Cai, Yunhe Zhang, Jicong Fan, and See-Kiong Ng. Lg-fgad: An effective federated graph anomaly detection framework. In Proceedings of the International Joint Conference on Artificial Intelligence , 2024

    2024

  3. [11]

    Hyperbolic graph convolutional neural networks

    Ines Chami, Zhitao Ying, Christopher R \'e , and Jure Leskovec. Hyperbolic graph convolutional neural networks. Advances in neural information processing systems , 32, 2019

    2019

  4. [12]

    Sampling techniques

    William Gemmell Cochran. Sampling techniques . john wiley & sons, 1977

    1977

  5. [13]

    Deep anomaly detection on attributed networks

    Kaize Ding, Jundong Li, Rohit Bhanushali, and Huan Liu. Deep anomaly detection on attributed networks. In Proceedings of the 2019 SIAM International Conference on Data Mining , pages 594--602, 2019

    2019

  6. [14]

    Uniform central limit theorems , volume 142

    Richard M Dudley. Uniform central limit theorems , volume 142. Cambridge university press, 2014

    2014

  7. [15]

    Graph mixture density networks

    Federico Errica, Davide Bacciu, and Alessio Micheli. Graph mixture density networks. ArXiv , abs/2012.03085, 2020

    2012 arXiv

  8. [16]

    Fast graph representation learning with pytorch geometric

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428 , 2019

    1903 arXiv

  9. [17]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning , pages 1263--1272, 2017

    2017

  10. [18]

    a tsch, Alexander J Smola, and Bernhard Sch \

    Arthur Gretton, Karsten M Borgwardt, Gunnar R \"a tsch, Alexander J Smola, and Bernhard Sch \"o lkopf. A kernel two-sample test. In Advances in neural information processing systems , pages 1299--1307, 2012

    2012

  11. [19]

    node2vec: Scalable feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages 855--864, 2016

    2016

  12. [20]

    Spectro-riemannian graph neural networks

    Karish Grover, Haiyang Yu, Xiang Song, Qi Zhu, Han Xie, Vassilis N Ioannidis, and Christos Faloutsos. Spectro-riemannian graph neural networks. arXiv preprint arXiv:2502.00401 , 2025

    2025 arXiv

  13. [21]

    Graphmore: Mitigating topological heterogeneity via mixture of riemannian experts

    Zihao Guo, Qingyun Sun, Haonan Yuan, Xingcheng Fu, Min Zhou, Yisen Gao, and Jianxin Li. Graphmore: Mitigating topological heterogeneity via mixture of riemannian experts. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 11754--11762, 2025

    2025

  14. [22]

    Exploring network structure, dynamics, and function using networkx

    Aric Hagberg, Pieter J Swart, and Daniel A Schult. Exploring network structure, dynamics, and function using networkx. Technical report, Los Alamos National Laboratory (LANL), Los Alamos, NM (United States), 2008

    2008

  15. [23]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in neural information processing systems , volume 30, 2017

    2017

  16. [24]

    Graph representation learning

    William L Hamilton. Graph representation learning . Morgan & Claypool Publishers, 2020

    2020

  17. [25]

    Stochastic blockmodels: First steps

    Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. Stochastic blockmodels: First steps. Social networks , 5(2):109--137, 1983

    1983

  18. [26]

    Anemone: Graph anomaly detection with multi-scale contrastive learning

    Ming Jin, Yixin Liu, Yu Zheng, Lianhua Chi, Yuan-Fang Li, and Shirui Pan. Anemone: Graph anomaly detection with multi-scale contrastive learning. In Proceedings of the 30th ACM international conference on information & knowledge management , pages 3122--3126, 2021

    2021

  19. [27]

    Wei Jin, Yao Ma, Yiqi Wang, Xiaorui Liu, Jiliang Tang, Yukuo Cen, Jie Tang, Chuan Shi, Yanfang Ye, Jiawei Zhang, and Philip S. Yu. Graph representation learning: Foundations, methods, applications and systems. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discover...

    2021

  20. [28]

    Methods of reducing sample size in monte carlo computations

    Herman Kahn and Andy W Marshall. Methods of reducing sample size in monte carlo computations. Journal of the Operations Research Society of America , 1(5):263--278, 1953

    1953

  21. [29]

    Advances and open problems in federated learning

    Peter Kairouz, H Brendan McMahan, Brendan Avent, Aur \'e lien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and Trends in Machine Learning , 1...

    2021

  22. [30]

    Marginalized kernels between labeled graphs

    Hisashi Kashima, Koji Tsuda, and Akihiro Inokuchi. Marginalized kernels between labeled graphs. In Proceedings of the 20th international conference on machine learning (ICML-03) , pages 321--328, 2003

    2003

  23. [31]

    Robust kernel density estimation

    JooSeuk Kim and Clayton D Scott. Robust kernel density estimation. The Journal of Machine Learning Research , 13(1):2529--2565, 2012

    2012

  24. [32]

    Rethinking reconstruction-based graph-level anomaly detection: limitations and a simple remedy

    Sunwoo Kim, Soo Yong Lee, Fanchen Bu, Shinhwan Kang, Kyungho Kim, Jaemin Yoo, and Kijung Shin. Rethinking reconstruction-based graph-level anomaly detection: limitations and a simple remedy. Advances in Neural Information Processing Systems , 37:95931--95962, 2024

    2024

  25. [33]

    Variational graph auto-encoders

    Thomas N Kipf and Max Welling. Variational graph auto-encoders. arXiv preprint arXiv:1611.07308 , 2016

    2016 arXiv

  26. [34]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations , 2017

    2017

  27. [35]

    Explainable classification of brain networks via contrast subgraphs

    Tommaso Lanciano, Francesco Bonchi, and Aristides Gionis. Explainable classification of brain networks via contrast subgraphs. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 3308--3318, 2020

    2020

  28. [36]

    Graphde: A generative framework for debiased learning and out-of-distribution detection on graphs

    Zenan Li, Qitian Wu, Fan Nie, and Junchi Yan. Graphde: A generative framework for debiased learning and out-of-distribution detection on graphs. Advances in Neural Information Processing Systems , 35:30277--30290, 2022

    2022

  29. [37]

    Cvtgad: Simplified transformer with cross-view attention for unsupervised graph-level anomaly detection

    Jindong Li, Qianli Xing, Qi Wang, and Yi Chang. Cvtgad: Simplified transformer with cross-view attention for unsupervised graph-level anomaly detection. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages 185--200. Springer, 2023

    2023

  30. [38]

    Isolation forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining , pages 413--422. IEEE, 2008

    2008

  31. [39]

    Graph normalizing flows

    Jenny Liu, Aviral Kumar, Jimmy Ba, Jamie Kiros, and Kevin Swersky. Graph normalizing flows. In Advances in neural information processing systems , pages 13578--13588, 2019

    2019

  32. [40]

    Energy-based models for atomic-resolution protein conformations

    Bowen Liu, Yutao Liu, Bozitao Liu, and Xiaolin Wang. Energy-based models for atomic-resolution protein conformations. In International Conference on Learning Representations , 2020

    2020

  33. [41]

    Good-d: On unsupervised graph out-of-distribution detection

    Yixin Liu, Kaize Ding, Huan Liu, and Shirui Pan. Good-d: On unsupervised graph out-of-distribution detection. In Proceedings of the 16th ACM International Conference on Web Search and Data Mining , pages 339--347, 2023

    2023

  34. [42]

    Towards self-interpretable graph-level anomaly detection

    Yixin Liu, Kaize Ding, Qinghua Lu, Fuyi Li, Leo Yu Zhang, and Shirui Pan. Towards self-interpretable graph-level anomaly detection. Advances in Neural Information Processing Systems , 36, 2023

    2023

  35. [43]

    Deep graph level anomaly detection with contrastive learning

    Xuexiong Luo, Jia Wu, Jian Yang, Shan Xue, Hao Peng, Chuan Zhou, Hongyang Chen, Zhao Li, and Quan Z Sheng. Deep graph level anomaly detection with contrastive learning. Scientific Reports , 12(1):19867, 2022

    2022

  36. [44]

    A comprehensive survey on graph anomaly detection with deep learning

    Xiaoxiao Ma, Jia Wu, Shan Xue, Jian Yang, Chuan Zhou, Quan Z Sheng, Hui Xiong, and Leman Akoglu. A comprehensive survey on graph anomaly detection with deep learning. IEEE Transactions on Knowledge and Data Engineering , 2021

    2021

  37. [45]

    Deep graph-level anomaly detection by glocal knowledge distillation

    Rongrong Ma, Guansong Pang, Ling Chen, and Anton van den Hengel. Deep graph-level anomaly detection by glocal knowledge distillation. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining , pages 704--714, 2022

    2022

  38. [46]

    Tudataset: A collection of benchmark datasets for learning with graphs

    Christopher Morris, Nils M Kriege, Franka Bause, Kristian Kersting, Petra Mutzel, and Marion Neumann. Tudataset: A collection of benchmark datasets for learning with graphs. arXiv preprint arXiv:2007.08663 , 2020

    2007 arXiv

  39. [47]

    Biological network analysis with deep learning

    Giulia Muzio, Leslie O’Bray, and Karsten Borgwardt. Biological network analysis with deep learning. Briefings in Bioinformatics , 22:1515--1530, 2020

    2020

  40. [48]

    Nachman and D

    B. Nachman and D. Shih. Anomaly detection with density estimation. Physical Review D , 101:075042, 2020

    2020

  41. [49]

    Propagation kernels: efficient graph kernels from propagated information

    Marion Neumann, Roman Garnett, Christian Bauckhage, and Kristian Kersting. Propagation kernels: efficient graph kernels from propagated information. Machine Learning , 102:209--245, 2016

    2016

  42. [50]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR) , 54(2):1--38, 2021

    2021

  43. [51]

    On estimation of a probability density function and mode

    Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics , 33(3):1065--1076, 1962

    1962

  44. [52]

    Deepwalk: Online learning of social representations

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , pages 701--710, 2014

    2014

  45. [53]

    Deep graph anomaly detection: A survey and new perspectives

    Hezhe Qiao, Hanghang Tong, Bo An, Irwin King, Charu Aggarwal, and Guansong Pang. Deep graph anomaly detection: A survey and new perspectives. arXiv preprint arXiv:2409.09957 , 2024

    2024 arXiv

  46. [54]

    Raising the bar in graph-level anomaly detection

    Chen Qiu, Marius Kloft, Stephan Mandt, and Maja Rudolph. Raising the bar in graph-level anomaly detection. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence , pages 2196--2203, 2022

    2022

  47. [55]

    Rong, Tingyang Xu, Junzhou Huang, Wen bing Huang, Hong Cheng, Yao Ma, Yiqi Wang, Tyler Derr, Lingfei Wu, and Tengfei Ma

    Y. Rong, Tingyang Xu, Junzhou Huang, Wen bing Huang, Hong Cheng, Yao Ma, Yiqi Wang, Tyler Derr, Lingfei Wu, and Tengfei Ma. Deep graph learning: Foundations, advances and applications. Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mi...

    2020

  48. [56]

    Temporal graph networks for deep learning on dynamic graphs

    Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 , 2020

    2006 arXiv

  49. [57]

    Estimating the support of a high-dimensional distribution

    Bernhard Sch \"o lkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. Estimating the support of a high-dimensional distribution. Neural computation , 13(7):1443--1471, 2001

    2001

  50. [58]

    Optimizing ood detection in molecular graphs: A novel approach with diffusion models

    Xu Shen, Yili Wang, Kaixiong Zhou, Shirui Pan, and Xin Wang. Optimizing ood detection in molecular graphs: A novel approach with diffusion models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages 2640--2650, 2024

    2024

  51. [59]

    Weisfeiler-lehman graph kernels

    Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research , 12(Sep):2539--2561, 2011

    2011

  52. [60]

    Grakel: A graph kernel library in python, 2020

    Giannis Siglidis, Giannis Nikolentzos, Stratis Limnios, Christos Giatsidis, Konstantinos Skianis, and Michalis Vazirgiannis. Grakel: A graph kernel library in python, 2020

    2020

  53. [61]

    Uniform: Towards unified framework for anomaly detection on graphs

    Chuancheng Song, Xixun Lin, Hanyang Shen, Yanmin Shang, and Yanan Cao. Uniform: Towards unified framework for anomaly detection on graphs. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 12559--12567, 2025

    2025

  54. [62]

    Spectral sparsification of graphs

    Daniel A Spielman and Shang-Hua Teng. Spectral sparsification of graphs. SIAM Journal on Computing , 40(4):981--1025, 2011

    2011

  55. [63]

    Mmd graph kernel: Effective metric learning for graphs via maximum mean discrepancy

    Yan Sun and Jicong Fan. Mmd graph kernel: Effective metric learning for graphs via maximum mean discrepancy. In The Twelfth International Conference on Learning Representations , 2024

    2024

  56. [64]

    Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization

    Fan-Yun Sun, Jordan Hoffmann, Vikas Verma, and Jian Tang. Infograph: Unsupervised and semi-supervised graph-level representation learning via mutual information maximization. arXiv preprint arXiv:1908.01000 , 2019

    1908 arXiv

  57. [65]

    Graph convolutional networks for computational drug development and discovery

    Mengying Sun, Sendong Zhao, Coryandar Gilvary, Olivier Elemento, Jiayu Zhou, and Fei Wang. Graph convolutional networks for computational drug development and discovery. Briefings in Bioinformatics , 2019

    2019

  58. [66]

    Learning graph representation via graph entropy maximization

    Ziheng Sun, Xudong Wang, Chris Ding, and Jicong Fan. Learning graph representation via graph entropy maximization. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st Int...

    2024

  59. [67]

    Introduction to Nonparametric Estimation

    Alexandre B Tsybakov. Introduction to Nonparametric Estimation . Springer, New York, NY, 2009

    2009

  60. [68]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research , 9(11), 2008

    2008

  61. [69]

    Deep graph infomax

    Petar Velickovic, William Fedus, William L Hamilton, Pietro Li \`o , Yoshua Bengio, and R Devon Hjelm. Deep graph infomax. ICLR (Poster) , 2(3):4, 2019

    2019

  62. [70]

    Graph kernels

    S Vichy N Vishwanathan, Nicol N Schraudolph, Risi Kondor, and Karsten M Borgwardt. Graph kernels. Journal of Machine Learning Research , 11(Apr):1201--1242, 2010

    2010

  63. [71]

    Learning low-dimensional latent graph structures: A density estimation approach

    Li Wang and Ren-Cang Li. Learning low-dimensional latent graph structures: A density estimation approach. IEEE Transactions on Neural Networks and Learning Systems , 31:1098--1112, 2020

    2020

  64. [72]

    Relational graph attention network for aspect-based sentiment analysis

    Kai Wang, Weizhou Shen, Yunyi Yang, Xiaojun Quan, and Rui Wang. Relational graph attention network for aspect-based sentiment analysis. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Com...

    2020

  65. [73]

    Graph neural networks for molecules

    Yuyang Wang, Zijie Li, and Amir Farimani. Graph neural networks for molecules. ArXiv , abs/2209.05582, 2022

    2022 arXiv

  66. [74]

    Unifying unsupervised graph-level anomaly detection and out-of-distribution detection: A benchmark

    Yili Wang, Yixin Liu, Xu Shen, Chenyu Li, Kaize Ding, Rui Miao, Ying Wang, Shirui Pan, and Xin Wang. Unifying unsupervised graph-level anomaly detection and out-of-distribution detection: A benchmark. arXiv preprint arXiv:2406.15523 , 2024

    2024 arXiv

  67. [75]

    Adaptive riemannian graph neural networks

    Xudong Wang, Tongxin Li, Chris Ding, and Jicong Fan. Adaptive riemannian graph neural networks. arXiv preprint arXiv:2508.02600 , 2025

    2025

  68. [76]

    Explainable graph representation learning via graph pattern analysis

    Xudong Wang, Ziheng Sun, Chris Ding, and Jicong Fan. Explainable graph representation learning via graph pattern analysis. In James Kwok, editor, Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25 , pages 3426--3434. Internatio...

    2025

  69. [77]

    Deep graph library: Towards efficient and scalable deep learning on graphs

    Minjie Yu Wang. Deep graph library: Towards efficient and scalable deep learning on graphs. In ICLR workshop on representation learning on graphs and manifolds , 2019

    2019

  70. [78]

    All of nonparametric statistics

    Larry Wasserman. All of nonparametric statistics . Springer Science & Business Media, 2006

    2006

  71. [79]

    Collective dynamics of ‘small-world’networks

    Duncan J Watts and Steven H Strogatz. Collective dynamics of ‘small-world’networks. nature , 393(6684):440--442, 1998

    1998

  72. [80]

    Using the nystr \"o m method to speed up kernel machines

    Christopher Williams and Matthias Seeger. Using the nystr \"o m method to speed up kernel machines. Advances in neural information processing systems , 13, 2000

    2000

  73. [81]

    A comprehensive survey on graph neural networks

    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, 2020

    2020

  74. [82]

    Rethinking explaining graph neural networks via non-parametric subgraph matching

    Fang Wu, Siyuan Li, Xurui Jin, Yinghui Jiang, Dragomir Radev, Zhangming Niu, and Stan Z Li. Rethinking explaining graph neural networks via non-parametric subgraph matching. In International conference on machine learning , pages 37511--37523. PMLR, 2023

    2023

  75. [83]

    Federated graph classification over non-iid graphs

    Han Xie, Jing Ma, Li Xiong, and Carl Yang. Federated graph classification over non-iid graphs. Advances in Neural Information Processing Systems , 34:18839--18852, 2021

    2021

  76. [84]

    How powerful are graph neural networks? In International Conference on Learning Representations , 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations , 2019

    2019

  77. [85]

    Infogcl: Information-aware graph contrastive learning

    Dongkuan Xu, Wei Cheng, Dongsheng Luo, Haifeng Chen, and Xiang Zhang. Infogcl: Information-aware graph contrastive learning. Advances in Neural Information Processing Systems , 34:30414--30425, 2021

    2021

  78. [86]

    Gnnexplainer: Generating explanations for graph neural networks

    Zhitao Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems , 32, 2019

    2019

  79. [87]

    Graph contrastive learning with augmentations

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. Graph contrastive learning with augmentations. Advances in neural information processing systems , 33:5812--5823, 2020

    2020

  80. [88]

    Dual-discriminative graph neural network for imbalanced graph-level anomaly detection

    Ge Zhang, Zhenyu Yang, Jia Wu, Jian Yang, Shan Xue, Hao Peng, Jianlin Su, Chuan Zhou, Quan Z Sheng, Leman Akoglu, et al. Dual-discriminative graph neural network for imbalanced graph-level anomaly detection. Advances in Neural Information Processing Systems , 35:24144--24157, 2022

    2022

  81. [89]

    Using classification datasets to evaluate graph outlier detection: Peculiar observations and new insights

    Lingxiao Zhao and Leman Akoglu. Using classification datasets to evaluate graph outlier detection: Peculiar observations and new insights. Big Data , 2021

    2021

  82. [90]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.