{"id":"d85325e5-095f-4e68-80ab-e832208abf4a","arxiv_id":"1909.01084","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MEGAN combines a fusing generator with per-view connectivity generators and a node-pair discriminator to embed multi-view networks, outperforming five baselines on two datasets.","lead":"MEGAN is a generative adversarial network that learns a shared low-dimensional embedding of a network with multiple types of edges, called views. The authors report that it beats five baseline methods on node classification, link prediction, and visualization using the Flickr and Last.fm datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MEGAN's generator does not define a coherent sampling distribution: Eq. (2) conditions on (vi,vj) while Eq. (6) uses G(K_ic|vi,vc), and the policy gradient drops the normalization; the omitted proof of Proposition 1 cannot repair the undefined pg.","rationale":"The reader identified exactly the same soft spot: the generator's probability model over candidate fake nodes is inconsistent. I agree that this is the most load-bearing concern. The only mechanism by which the embedding X is learned is the policy-gradient update in Eq. (6), and that update is valid only if pg is a coherent, normalized distribution over negative nodes and if the sampled actions follow it. As written, Eq. (2) scores candidates with the fused representation of the positive pair, Eq. (6) uses G(K_ic|vi,vc) as a node distribution, and the deterministic argmax selection in Algorithm 1 is not a draw from pg. The omitted proof of Proposition 1 cannot repair this because the proposition's hypothesis ('pg converges to pdata') is not tied to a defined object. I do not see a second concern that is more central: the lack of error bars and code affects confidence in the empirical results, but the conditioning mismatch directly undermines the claimed learning procedure. The concern is fixable by defining pg, correcting Eq. (2), and re-deriving the gradient. Since the reader already issued CONDITIONAL, no verdict change is needed; the condition should explicitly require this corrected derivation and ideally a released implementation before the central claim is accepted.","tokens_in":10538,"tokens_out":6143,"duration_ms":62526,"concrete_test":"Re-derive Eq. (6) after defining pg(vc|vi)=G(K_ic|vi,vc)/Σ_{c'}G(K_ic'|vi,vc') and correcting Eq. (2) to condition on (vi,vc) rather than (vi,vj). If the resulting REINFORCE gradient contains an extra ∇θG log Σ_{c'} G(K_ic'|vi,vc') term or otherwise differs from Eq. (6), then Algorithm 1 does not minimize Eq. (4), and the central training mechanism is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that Algorithm 1 actually optimizes the adversarial objective in Eq. (4). That requires a well-defined distribution pg over fake pairs (vi,vc) that the generator can sample from and whose score-function gradient is given by Eq. (6). The paper never defines pg. Eq. (2) selects vc by arg max over product_l G^(l)(e_ic^(l)=e_ij^(l)|vi,vj,Gf(vi,vj;...)): the score is conditioned on the positive pair (vi,vj), not on the candidate pair (vi,vc). Eq. (6) then treats G(K_ic|vi,vc) as the probability of drawing vc, which is a different model. Moreover, G(K_ic|vi,vc) is the probability of a connectivity pattern, not a probability mass function over nodes; to use it as pg one must normalize over vc, and the REINFORCE gradient must include the log-normalizer term. The derivation in Eq. (6) drops that term and further assumes sampling from G, while Algorithm 1 line 4 samples deterministically by argmax. Thus the generator update is not the gradient of the stated objective, and Proposition 1, whose proof is omitted, cannot establish convergence of pg to pdata when pg is undefined. The reported empirical gains may survive a corrected formulation, but the mechanism claimed for learning X is not justified as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":10874,"tokens_out":5805,"duration_ms":64993,"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":[{"comment":"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":"Section 3.2, Eqs. (2) and (6)"},{"comment":"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":"Section 3.4, Proposition 1"},{"comment":"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.","section":"Section 4 (Experiments)"}],"minor_comments":[{"comment":"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.","section":"Algorithm 1"},{"comment":"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":"Eq. (2)"},{"comment":"The description of DRNE contains a typo: 'constructs utilizes an LSTM' should be 'constructs and utilizes an LSTM' or similar.","section":"Section 4.2"},{"comment":"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.","section":"Section 4.3, link prediction"},{"comment":"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.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The main technical flaw is substantive but potentially fixable: the authors need to define a coherent generator distribution over candidate nodes, derive the correct policy-gradient update, and re-run the experiments after the correction. I would also require code and full experimental details before accepting. The omitted proof of Proposition 1 should be supplied or the convergence claim should be removed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: MEGAN is a reasonable extension of GraphGAN to multi-view networks, and the reported results are consistent wins on two datasets. But the generator's sampling distribution is not well-defined, and the policy gradient update doesn't follow from the stated objective. That's a load-bearing flaw, not a cosmetic one.\n\nWhat's new: the architecture combines a fusing generator that produces a common representation for a node pair, per-view connectivity generators, and a node-pair discriminator. That's a genuine extension of GraphGAN and MNE, and the paper cites the relevant prior work. The idea of generating fake pairs whose connectivity pattern matches the positive pair is sensible.\n\nThe experiments compare against node2vec, GraphGAN, DRNE, MVE, and MNE on Last.fm and Flickr. MEGAN wins on node classification, link prediction, and the visualization figures look plausible. No code, no error bars, but the gains are consistent across settings, so I believe the empirical claim may survive a corrected formulation.\n\nThe soft spot is in the derivation. Eq. (2) selects the fake node vc by maximizing a probability that is conditioned on the positive pair (vi,vj), not on the candidate pair (vi,vc). The product of G^(l) terms scores the event that the connectivity between vi and vc equals that of vi and vj, but the conditioning variable is (vi,vj). So the argmax over vc is not a probability distribution over nodes. Eq. (6) then writes the expectation over vc as G(Kic|vi,vc) times the log-discriminator term, treating G(Kic|vi,vc) as the probability of drawing vc. That is a different object: G(Kic|vi,vc) is the probability of a connectivity pattern, not a pmf over nodes. To use it as pg you would need to normalize over all vc, and the REINFORCE gradient in Eq. (6) drops that normalizer. The paper also omits the proof of Proposition 1, so the convergence claim can't rescue the gap.\n\nIn short, the paper is an honest attempt with a genuine architectural idea, but the generator as written has no coherent sampling distribution. The empirical results may survive a rewrite, but the mechanism claimed for learning X is not justified. For a reader who wants a GAN-based multi-view embedding, this is a useful starting point but not a reliable method as it stands. The paper deserves a serious referee: the idea is worth engaging with and the flaws are fixable, but it needs major revision, especially a corrected generator formulation, a real proof or removed claim, and released code with standard deviations.","headline":"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.","tokens_in":11396,"tokens_out":3159,"would_cite":false,"duration_ms":27275,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["multi-view network embedding","generative adversarial network","node classification","link prediction","network representation learning","adversarial sampling","policy gradient","multi-view network"],"falsifier":"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.","tokens_in":10353,"feed_emoji":"🕸️","tokens_out":9987,"duration_ms":90284,"temperature":0.7,"pith_summary":"This paper introduces MEGAN, a generative adversarial network for multi-view network embedding—learning one low-dimensional vector per node when the same nodes are connected by several different kinds of relations. The authors aim to show that an adversarial game between a generator that fabricates node pairs and a discriminator that labels pairs real or fake can capture the connectivity of each view as well as the correlations between views. If this works, it gives a general unsupervised way to create features for node classification, link prediction, and visualization on multi-view data, where most prior methods are either single-view or use simpler fusion rules. The paper reports experiments on Last.fm and Flickr in which MEGAN outperforms five existing single-view and multi-view embedding methods on all three tasks.","feed_headline":"Fusing network views into one embedding outperforms single-view models","feed_subtitle":"Adversarial training preserves link structure within each view and correlations across views.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"It defines the GAN minmax objective and the convergence argument that MEGAN's Proposition 1 claims to extend to discrete multi-view data.","marker":"[Goodfellow et al., 2014]"},{"why":"GraphGAN is the single-view GAN embedding method MEGAN builds on and the baseline it improves, and it also supplies the policy-gradient update for discrete node samples.","marker":"[Wang et al., 2018]"},{"why":"MVE is the multi-view embedding baseline that fuses per-view embeddings by weighted combination and must be outperformed.","marker":"[Qu et al., 2017]"},{"why":"MNE is the multi-view embedding baseline that integrates per-view embeddings through a shared latent space and must be outperformed.","marker":"[Zhang et al., 2018a]"},{"why":"node2vec is a single-view random-walk baseline and is the source of the link-prediction evaluation protocol used in the experiments.","marker":"[Grover and Leskovec, 2016]"},{"why":"DRNE is the single-view baseline that recursively aggregates node neighborhoods with an LSTM.","marker":"[Tu et al., 2018]"},{"why":"It provides the policy-gradient training technique that MEGAN uses to update the generator over discrete node choices.","marker":"[Yu et al., 2017]"},{"why":"It supplies the Last.fm and Flickr multi-view network data sets on which all comparisons are run.","marker":"[Bui et al., 2016]"}],"fun_headline_variants":["GAN-based multi-view network embedding beats single-view and prior methods","MEGAN: adversarial training fuses multiple network views into one embedding","Multi-view network embedding via GAN: better links, nodes, and visuals","MEGAN: connecting views with adversarial learning for superior embeddings","Harnessing GANs for multi-view networks: MEGAN outperforms on three tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["GAN-based multi-view network embedding beats single-view and prior methods","MEGAN: adversarial training fuses multiple network views into one embedding","Multi-view network embedding via GAN: better links, nodes, and visuals","MEGAN: connecting views with adversarial learning for superior embeddings","Harnessing GANs for multi-view networks: MEGAN outperforms on three tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00073,"raw_usage":{"total_tokens":3237,"prompt_tokens":885,"completion_tokens":2352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":2254}},"tokens_in":501,"tokens_out":2352,"duration_ms":16461,"temperature":1.0,"reasoning_tokens":2254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:25:13.353371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Graphgan: graph representation learning with generative adversarial nets","cited_arxiv_id":null,"evidence_quote":"GraphGAN is the single-view GAN embedding method MEGAN builds on and the baseline it improves, and it also supplies the policy-gradient update for discrete node samples."},{"cited_title":"node2vec: Scalable feature learning for networks","cited_arxiv_id":null,"evidence_quote":"node2vec is a single-view random-walk baseline and is the source of the link-prediction evaluation protocol used in the experiments."},{"cited_title":"Labeling actors in multi-view social networks by integrat- ing information from within and across multiple views","cited_arxiv_id":null,"evidence_quote":"It supplies the Last.fm and Flickr multi-view network data sets on which all comparisons are run."}],"review_version":1}