REVIEW 3 major objections 5 minor 25 references
MEGAN: A Generative Adversarial Network for Multi-View Network Embedding
T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read MEGAN learns a single multi-view network embedding by pitting a fake-pair generator against a node-pair discriminator, and the authors report it beats five baselines on two real-world datasets.
desk verdict MEGAN's architecture is a genuine extension of GraphGAN to multi-view networks, but the generator's sampling distribution is ill-defined and the policy gradient update does not follow from the stated objective. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the multi-view generator: a fusing generator $G_f$ that merges the embeddings of a node pair into one representation, and $k$ connectivity generators $G^{(l)}$ that convert that fused representation into a per-view probability that $e^{(l)}_{ij}=1$, combined as a product in Eq. (1). A node-pair discriminator $D(v_i,v_j) = \mathrm{sigmoid}(d_i^\top d_j)$ separates real from fake pairs. The training loop alternates discriminator updates with a policy-gradient generator update, which routes the discrete choice of fake node back into the embedding matrix $X$; this is how the embedding is forced to encode per-view connectivity and inter-view correlation.
What would settle it
On a small multi-view graph, enumerate all candidate nodes and compare the probability of choosing each fake node under Eq. (2) with the probability assigned by the generator model in Eq. (6); if the two disagree, the training loop is not optimizing the stated minmax objective and the claimed convergence does not follow from the update rule.
Extended reading notes
Core claim
The central claim is that when the adversarial game reaches equilibrium, the learned embedding matrix $X$ captures both the connectivity within each view and the correlations across views, so that the generator's distribution over fake node pairs approximates the true multi-view connectivity. For a real pair $(v_i, v_j)$, the generator selects a fake partner $v_c$ whose connectivity pattern $K_{ic}$ matches $K_{ij}$ as closely as possible, and the discriminator learns to tell such fake pairs from real ones. The paper asserts that this training produces embeddings that outperform MVE, MNE, node2vec, GraphGAN, and DRNE on node classification, link prediction, and visualization on the Last.fm and Flickr datasets.
Load-bearing premise
The load-bearing premise is that the generator's discrete node selection actually maximizes the adversarial objective, i.e., that the score used to pick a fake node in Eq. (2) and the probability model assumed by the policy-gradient update in Eq. (6) describe the same distribution.
Editorial extensions
If this is right
- On any multi-view network, the embedding can be learned without labels and then plugged into standard classifiers or scoring functions for node classification and link prediction.
- Because the $k$ connectivity generators factorize given the fused representation, the framework scales to an arbitrary number of views and each view's generator can be trained in parallel.
- The performance gap over single-view methods should be largest when the views carry complementary rather than redundant information, which is the explanation the paper gives for its link-prediction results.
- The generator's factorized model provides an explicit probability for a node pair's connectivity pattern, so the learned representation has a direct probabilistic interpretation rather than being a purely black-box embedding.
Reading between the lines
- A null-data experiment could shuffle edges across views to destroy cross-view correlation; if MEGAN still beats the single-view baselines, the gain would come from per-view connectivity rather than the claimed correlation capture.
- A reader could enumerate all candidate nodes on a small graph and check whether the distribution used to select fake nodes in the sampling step matches the distribution implied by the policy-gradient update; if the two differ, the implemented training loop would not be optimizing the objective stated in Eq. (4).
- The same adversarial node-selection scheme could in principle be applied to dynamic multi-view networks, where the discrete action set changes over time; the paper does not explore that setting.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes MEGAN, a generative adversarial network for multi-view network embedding. The method consists of a generator that produces fake node pairs by selecting a candidate node whose multi-view connectivity pattern matches that of a real pair, and a discriminator that distinguishes real from fake node pairs. The generator is composed of a fusing component and per-view connectivity generators, and the learned embedding matrix X is claimed to capture within-view connectivity and cross-view correlations after the adversarial game converges. The authors present a min-max objective, a policy-gradient update for the generator, a convergence proposition, and experiments on Last.fm and Flickr comparing MEGAN with node2vec, GraphGAN, DRNE, MVE, and MNE on node classification, link prediction, and visualization, reporting that MEGAN outperforms the baselines.
Significance. If the method and experiments were fully substantiated, MEGAN would be a useful contribution to multi-view network representation learning: the idea of adversarially generating fake node pairs whose connectivity patterns are similar across views is a natural and timely extension of GraphGAN, and the empirical task setup follows common practice in the field. However, the manuscript as written does not establish the central claims. The generator distribution is not coherently defined, the convergence proof is omitted, the reported experiments lack code, variance estimates, and full hyperparameter details, and the superiority claim rests on only two data sets. The conceptual direction is promising, but the technical and empirical support is currently insufficient for publication.
major comments (3)
- [Section 3.2, Eqs. (2) and (6)] The generator sampling distribution pg is never defined, and the two equations that characterize it are inconsistent. Eq. (2) selects a negative node by maximizing a product of per-view probabilities conditioned on the positive pair (vi,vj) and on the fused representation Gf(vi,vj; X, θf), whereas Eq. (6) treats G(Kic|vi,vc) as the probability of drawing vc from the generator. These are different probability models: one conditions on the positive pair, the other on the candidate pair. Moreover, G(Kic|vi,vc) is a probability over connectivity patterns, not a probability mass function over candidate nodes, so using it as an unnormalized pg requires a partition function over V; the REINFORCE gradient in Eq. (6) drops the corresponding log-normalizer term. In addition, Algorithm 1 line 4 samples negative nodes by argmax selection, not by sampling from any stated pg. Consequently, the paper does not demonstrate that Algorithm 1 optimizes the objective in Eq. (4), and the claimed mechanism for learning X is not justified as written.
- [Section 3.4, Proposition 1] Proposition 1 is stated as the theoretical basis for convergence of pg to pdata, but its proof is explicitly omitted with the comment that it is similar to Goodfellow et al. This is not a routine adaptation: the setting involves discrete node selection, an argmax-based sampling procedure, a policy-gradient update, and an undefined generator distribution. Even a standard GAN convergence argument would require a well-defined pg and a proof that the update rule is the gradient of the stated objective; neither condition is established. The proposition should either be proved under explicit assumptions or replaced by a clearly qualified claim.
- [Section 4 (Experiments)] The central claim is empirical, but the experimental report is not reproducible or statistically substantiated. No code or implementation details are provided; Fig. 3 reports values averaged over 10 runs but shows no error bars or standard deviations; Fig. 4 does not report repeated trials or variance; and no significance tests are conducted. In addition, the hyperparameters for MEGAN are not fully specified beyond the embedding dimension and the general structure of Algorithm 1 (e.g., learning rates, numbers of generator and discriminator steps per iteration, minibatch sizes, and the values of s and t are not given), so the comparison with baselines cannot be independently verified or assessed for fairness.
minor comments (5)
- [Algorithm 1] The algorithm is titled 'MVGAN framework' although the method is called MEGAN, and line 5 refers to updating according to Eq. (1) and Eq. (6), but Eq. (1) is the generative model, not an update rule.
- [Eq. (2)] The displayed formula appears to be missing θ^(l) and a closing parenthesis in the argument of G^(l); it should read G^(l)(e^(l)_ic = e^(l)_ij | vi, vj, Gf(vi, vj; X, θf); θ^(l)).
- [Section 4.2] The description of DRNE contains a typo: 'constructs utilizes an LSTM' should be 'constructs and utilizes an LSTM' or similar.
- [Section 4.3, link prediction] The link prediction setup removes 50% of the edges in one view, but the construction of negative examples is not fully specified; it is unclear whether negative edges are pairs absent in all views or absent only in the target view, which affects the interpretation of the AUC/AP results.
- [Figure 6] The dimension-sensitivity analysis reports only MEGAN's performance; adding the best-performing baselines to the same figure would help the reader assess whether the observed trend is specific to MEGAN.
Circularity Check
No significant circularity: MEGAN's empirical claims rest on external baselines; the generator inconsistency is a correctness gap, not a circular reduction.
full rationale
The paper's central derivation is the min-max objective in Eq. (4), with the generator defined by Eq. (1), the negative-sampling rule in Eq. (2), and training updates in Eqs. (5)-(6). The claim that the learned embedding X captures intra-view connectivity and inter-view correlations is a design goal of that objective rather than a quantity fitted to the evaluation labels; node classification and link prediction are evaluated with an external classifier on embeddings learned without labels (Section 4.2). The theoretical convergence assertion (Proposition 1) is explicitly imported from Goodfellow et al. 2014 with the proof omitted; an omitted proof is an evidentiary gap, not a circular reduction. The only self-citation of note is the dataset attribution to Bui et al. 2016, which includes co-author Honavar, but the datasets are independent benchmark collections and are not used to define the algorithm. The inconsistency between Eq. (2), which conditions on the positive pair (vi,vj), and Eq. (6), which treats G(Kic|vi,vc) as a sampling distribution, is a correctness concern about the generator update, but it does not make the empirical comparison circular. No fitted parameter is renamed as a prediction, and no load-bearing result reduces to the paper's own definitions or self-citations.
Assumptions & free parameters
free parameters (3)
- embedding dimension d =
128
- discriminator sample sizes s and t
- training hyperparameters (learning rate, epochs, batch size, pretraining)
assumptions (3)
- ad hoc to paper Convergence of the adversarial game to pdata in the discrete multi-view setting (Proposition 1)
- domain assumption Conditional independence of views given the fused representation
- domain assumption Negative candidates can be restricted to the neighbor set N(vi)
Cite this review
Pith. "Pith review of MEGAN: A Generative Adversarial Network for Multi-View Network Embedding." pith.science (2026). https://pith.science/paper/IILT3NCF
@misc{pith2026190901084,
author = {Pith},
title = {Pith review of: MEGAN: A Generative Adversarial Network for Multi-View Network Embedding},
year = {2026},
howpublished = {\url{https://pith.science/paper/IILT3NCF}},
note = {Machine review of arXiv:1909.01084}
}
read the original abstract
Data from many real-world applications can be naturally represented by multi-view networks where the different views encode different types of relationships (e.g., friendship, shared interests in music, etc.) between real-world individuals or entities. There is an urgent need for methods to obtain low-dimensional, information preserving and typically nonlinear embeddings of such multi-view networks. However, most of the work on multi-view learning focuses on data that lack a network structure, and most of the work on network embeddings has focused primarily on single-view networks. Against this background, we consider the multi-view network representation learning problem, i.e., the problem of constructing low-dimensional information preserving embeddings of multi-view networks. Specifically, we investigate a novel Generative Adversarial Network (GAN) framework for Multi-View Network Embedding, namely MEGAN, aimed at preserving the information from the individual network views, while accounting for connectivity across (and hence complementarity of and correlations between) different views. The results of our experiments on two real-world multi-view data sets show that the embeddings obtained using MEGAN outperform the state-of-the-art methods on node classification, link prediction and visualization tasks.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Laplacian eigenmaps and spectral techniques for embedding and clustering
[Belkin and Niyogi, 2001] Mikhail Belkin and Partha Niyogi. Laplacian eigenmaps and spectral techniques for embedding and clustering. In NIPS, volume 14, pages 585–591,
work page 2001
-
[7]
Inductive representation learning on large graphs
[Hamilton et al., 2017] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pages 1024–1034,
work page 2017
-
[8]
Detecting overlapping commu- nities from local spectral subspaces
[He et al., 2015] Kun He, Yiwei Sun, David Bindel, John Hopcroft, and Yixuan Li. Detecting overlapping commu- nities from local spectral subspaces. In ICDM, pages 769–
work page 2015
-
[10]
[Kivel¨a et al., 2014] Mikko Kivel ¨a, Alex Arenas, Marc Barthelemy, James P Gleeson, Yamir Moreno, and Ma- son A Porter. Multilayer networks. Journal of complex networks, 2(3):203–271,
work page 2014
-
[11]
Multi-view clustering with graph embedding for connectome analysis
[Ma et al., 2017] Guixiang Ma, Lifang He, Chun-Ta Lu, Weixiang Shao, Philip S Yu, Alex D Leow, and Ann B Ragin. Multi-view clustering with graph embedding for connectome analysis. In CIKM. ACM,
work page 2017
-
[12]
Multi-dimensional graph convolutional networks
[Ma et al., 2019] Yao Ma, Suhang Wang, Chara C Aggarwal, Dawei Yin, and Jiliang Tang. Multi-dimensional graph convolutional networks. In SDM, pages 657–665. SIAM,
work page 2019
-
[15]
Deepwalk: Online learning of social rep- resentations
[Perozzi et al., 2014] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social rep- resentations. In Proceedings of SIGKDD, pages 701–710. ACM,
work page 2014
-
[16]
An attention-based collaboration framework for multi-view network represen- tation learning
[Qu et al., 2017] Meng Qu, Jian Tang, Jingbo Shang, Xiang Ren, Ming Zhang, and Jiawei Han. An attention-based collaboration framework for multi-view network represen- tation learning. In CIKM, pages 1767–1776. ACM,
work page 2017
Show all 25 references
-
[17]
Dynamics of large multi-view social networks: Synergy, cannibalization and cross-view interplay
[Shi et al., 2016] Yu Shi, Myunghwan Kim, Shaunak Chat- terjee, Mitul Tiwari, Souvik Ghosh, and R ´omer Rosales. Dynamics of large multi-view social networks: Synergy, cannibalization and cross-view interplay. In KDD, pages 1855–1864. ACM,
2016
-
[18]
Line: Large- scale information network embedding
[Tang et al., 2015] Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large- scale information network embedding. In Proceedings of WWW, pages 1067–1077,
2015
-
[19]
Visualizing large-scale and high- dimensional data
[Tang et al., 2016] Jian Tang, Jingzhou Liu, Ming Zhang, and Qiaozhu Mei. Visualizing large-scale and high- dimensional data. In Proceedings of WWW , pages 287– 297,
2016
-
[20]
Deep recursive network embedding with regular equivalence
[Tu et al., 2018] Ke Tu, Peng Cui, Xiao Wang, Philip S Yu, and Wenwu Zhu. Deep recursive network embedding with regular equivalence. In Proceedings of SIGKDD , pages 2357–2366. ACM,
2018
-
[21]
On deep multi-view represen- tation learning
[Wang et al., 2015] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. On deep multi-view represen- tation learning. In International Conference on Machine Learning, pages 1083–1092,
2015
-
[22]
Graphgan: graph representation learning with generative adversarial nets
[Wang et al., 2018] Hongwei Wang, Jia Wang, Jialin Wang, Miao Zhao, Weinan Zhang, Fuzheng Zhang, Xing Xie, and Minyi Guo. Graphgan: graph representation learning with generative adversarial nets. In AAAI,
2018
-
[23]
Seqgan: Sequence generative adversarial nets with policy gradient
[Yu et al., 2017] Lantao Yu, Weinan Zhang, Jun Wang, and Yong Yu. Seqgan: Sequence generative adversarial nets with policy gradient. In AAAI,
2017
-
[24]
Representation learning for large-scale dynamic networks
[Yu et al., 2018] Yanwei Yu, Huaxiu Yao, Hongjian Wang, Xianfeng Tang, and Zhenhui Li. Representation learning for large-scale dynamic networks. In International Con- ference on Database Systems for Advanced Applications , pages 526–541. Springer,
2018
-
[25]
Arbitrary-order proximity preserved network embedding
[Zhang et al., 2018b] Ziwei Zhang, Peng Cui, Xiao Wang, Jian Pei, Xuanrong Yao, and Wenwu Zhu. Arbitrary-order proximity preserved network embedding. In SIGKDD, pages 2778–2786, 2018
2018
-
[2001]
Net- gan: Generating graphs via random walks
[Bojchevski et al., 2018] Aleksandar Bojchevski, Oleksandr Shchur, Daniel Z ¨ugner, and Stephan G ¨unnemann. Net- gan: Generating graphs via random walks. arXiv preprint arXiv:1803.00816,
2018 arXiv
-
[2008]
Asymmetric transitivity preserv- ing graph embedding
[Ou et al., 2016] Mingdong Ou, Peng Cui, Jian Pei, Ziwei Zhang, and Wenwu Zhu. Asymmetric transitivity preserv- ing graph embedding. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discov- ery and data mining , pages 1105–1114. ACM,
2016
-
[2014]
node2vec: Scalable feature learning for networks
[Grover and Leskovec, 2016] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of SIGKDD , pages 855–864. ACM,
2016
-
[2015]
Zheng, Wen Wen, Yifan Yang, and Yuanmi Chen
[Huang et al., 2018] Jiaming Huang, Zhao Li, Vincent W. Zheng, Wen Wen, Yifan Yang, and Yuanmi Chen. Unsu- pervised multi-view nonlinear graph embedding. In UAI, pages 319–328,
2018
-
[2016]
Learning community embedding with community de- tection and node embedding on graphs
[Cavallari et al., 2017] Sandro Cavallari, Vincent W Zheng, Hongyun Cai, Kevin Chen-Chuan Chang, and Erik Cam- bria. Learning community embedding with community de- tection and node embedding on graphs. In CIKM, pages 377–386,
2017
-
[2017]
Gen- erative adversarial nets
[Goodfellow et al., 2014] Ian Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial nets. In NIPS, pages 2672–2680,
2014
-
[2018]
Labeling actors in multi-view social networks by integrat- ing information from within and across multiple views
[Bui et al., 2016] Ngot Bui, Thanh Le, and Vasant Honavar. Labeling actors in multi-view social networks by integrat- ing information from within and across multiple views. In BigData, pages 616–625. IEEE,
2016
-
[2019]
Visualizing data using t-sne
[Maaten and Hinton, 2008] Laurens van der Maaten and Ge- offrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579–2605,
2008
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.