{"id":"2c200491-340e-46a0-966e-e7fae3bd8fc6","arxiv_id":"2412.15582","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"DG-Gen directly models the probability of temporal edges as a product of conditional distributions and autoregressively generates continuous-time dynamic graphs with node and edge features.","lead":"DG-Gen is a new generative model that treats each event in a continuously evolving network as a joint probability of source, destination, time, and edge features, sampled autoregressively. The authors show it produces synthetic graphs with similar statistics to real interaction data and competitive link prediction scores.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inductive-generation claim is unsupported: p(src) and p(dst|src) are categorical over a fixed node vocabulary, and the Section 3.4 generation loop starts from an empty graph with no node-arrival mechanism, so unseen nodes cannot actually be synthesized.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: the categorical node distributions in Section 3.2 cannot support inductive generation of unseen nodes because they sample from a fixed, pre-existing set of node embeddings, and Section 3.4 provides no node-creation mechanism to expand that set. This is not a stylistic or incremental issue; the paper's central novelty, repeated in the abstract, Section 1, and Table 1, is that DG-Gen is 'fully inductive by default' and 'generates graphs with unseen nodes and timestamps.' If this claim fails, the method reduces to a transductive edge-sequence generator, which is not the contribution being argued. The contradiction in Section 3.4—starting with an empty graph yet having 'all nodes' with empty memory—highlights that the node vocabulary must be supplied externally, which is transductive. The experimental sections do not settle this because they only compare aggregate graph statistics and edge-overlap, not whether any generated node is novel. Other concerns, such as the Exponential inter-event time assumption and imported baselines without error bars, are secondary: they affect robustness and comparison fairness, but the inductive claim is the load-bearing pillar. The proposed test directly checks whether the released code can generate from an empty node set or whether all generated nodes come from the training set, which would definitively confirm or refute the concern. Because the reader already rejected the paper on this basis, the verdict should remain unchanged.","tokens_in":17339,"tokens_out":3467,"duration_ms":31600,"concrete_test":"Run the released generation code (github.com/ryienh/DGGen) on Wikipedia with an empty initial node set and no pre-specified node IDs, following the Section 3.4 description; if the sampling loop cannot produce a first event or must be seeded with existing IDs, inductive generation is not implemented. As a complementary statistical check, train on a node-disjoint 90/10 split, generate a synthetic graph, and count the fraction of generated node IDs absent from the training node set; if this fraction is zero, the model never synthesizes unseen nodes. Either result settles whether the 'inductive' claim is real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is inductive CTDG generation (Section 1, Table 1): 'generates graphs with unseen nodes and timestamps.' This is not supported by the described model. In Section 3.2, p(src) is a Categorical distribution over 'possible source nodes' and p(dst|src) over 'possible destination nodes', with logits produced by Reshape/Product modules applied to temporal embeddings. These distributions have finite support: the set of node IDs with embeddings. Section 3.4 states 'DG-Gen starts with an empty graph and all nodes have empty memory'—but if the graph is empty, there are no nodes forming the support of p(src), so the first event cannot be sampled. To generate an unseen node, the process would need an explicit node-creation event that introduces a new ID, initializes its embedding, and expands the categorical support; no such mechanism is specified. If the implementation instead pre-populates the node set with training/test IDs, then 'unseen' nodes are not generated; they are provided as inputs, which is transductive. The experiments generate synthetic graphs with 'the same number of interactions as the test partition' but never report whether any generated node ID is outside the training set. Thus the central inductive claim is not established; at best the method is a transductive continuous-time edge generator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DG-Gen, a continuous-time dynamic graph (CTDG) generative framework that factorizes each temporal interaction as p(src, dst, t, e) = p(src)p(dst|src)p(t, e|dst, src). The encoder is a Temporal Graph Network, and the decoder uses learned modules to parameterize a categorical distribution over source nodes, a categorical distribution over destination nodes given the source, an exponential distribution for inter-event times, and categorical or Gaussian-mixture distributions for edge features. The authors claim that this factorization enables scalable, inductive, and largely assumption-free autoregressive generation of CTDGs, and they evaluate the method on five datasets for graph generation and link prediction, comparing against TIGGER-I and several discriminative baselines.","tokens_in":17616,"tokens_out":4198,"duration_ms":41710,"significance":"If the inductive claim were established, DG-Gen would be a conceptually clean departure from snapshot- or random-walk-based CTDG generators and, to my knowledge, the first CTDG generator that produces arbitrary-length edge features. The paper also deserves credit for releasing source code, including ablations of the memory and noise components, and for evaluating on five datasets. However, the paper's central claim—inductive generation of graphs with unseen nodes—is not supported by the model as described, and the temporal model rests on an untested exponential-assumption that is in tension with the 'largely assumption free' framing. These are load-bearing issues for the paper's main contribution.","major_comments":[{"comment":"The inductive generation claim is not supported by the described model. In §3.2, p(src) and p(dst|src) are Categorical distributions over 'possible source nodes' and 'possible destination nodes', with logits built from node embeddings; these distributions have finite support determined by the set of node IDs present in the embedding table. Section 3.4 states that generation starts with an empty graph and that all nodes have empty memory, but no node-arrival or node-creation mechanism is specified that would introduce new node IDs, initialize their embeddings, and expand the categorical support. Therefore the generative loop cannot sample an unseen node: at the first step there are no nodes in the support of p(src), and at later steps only nodes whose IDs are already in the support can appear. The experiments in §4 generate graphs with the same number of interactions as the test partition but never report whether any generated node ID is outside the training node set. As written, the method is a transductive edge-sequence generator, not an inductive node-generating model, and the headline contribution in §1 and Table 1 is unsubstantiated.","section":"§3.2 and §3.4"},{"comment":"The paper imposes, without empirical justification, that inter-event times follow an Exponential distribution conditioned on source and destination embeddings. This is a strong parametric assumption and directly contradicts the abstract's and §1's claim that the approach is 'largely assumption free'. Since the generated timestamps are one of the three components of the generated event tuple, the exponential assumption is load-bearing for temporal fidelity, yet no goodness-of-fit test, comparison with alternative inter-event distributions (e.g., Weibull, log-normal, or a nonparametric model), or residual analysis is reported. The absence of any evaluation of generated inter-event time statistics, as opposed to snapshot-based topological statistics in Table 2, makes it impossible to assess how well the temporal dynamics are reproduced.","section":"§3.2, p(t|dst,src)"},{"comment":"The link prediction results in Figure 3 are used to support the claim of 'significantly advances link prediction tasks', but the comparison may not be fully controlled. For all datasets except Bikeshare, baseline numbers are imported from Poursafaei et al. 2022 rather than re-run, and the manuscript does not state that the exact data splits, negative sampling strategy, and evaluation protocol are identical to the ones used for DG-Gen. In addition, the paper does not describe how p(dst|src) is applied when the candidate destination set includes nodes unseen during training, which is relevant because the claimed inductive link prediction setting depends on the model being able to score such destinations. These omissions weaken the secondary contribution, even though the primary issue is the generative claim.","section":"§4, link prediction comparison"}],"minor_comments":[{"comment":"The phrase 'largely assumption free' is too strong given the explicit Exponential distribution for inter-event times described in §3.2; the wording should be qualified.","section":"Abstract and §1"},{"comment":"There is a typo: 'continously' should be 'continuously'.","section":"§2.3"},{"comment":"The table header formatting is broken ('Capability DymondTagGenTIGGERTIGGER-IDG-Gen'), making the header difficult to read; also, the column alignment should be fixed.","section":"Table 1"},{"comment":"The caption says 'Best model is bolded', but in the manuscript text the bold formatting is not visible; please ensure the final PDF displays bolding clearly.","section":"Table 2"},{"comment":"The training description says a random subset of nodes is sampled when computing source and destination scores, but it is not stated whether the softmax normalization is restricted to that subset or applied over the full node set; this affects the interpretation of the reported negative log-likelihood.","section":"§3.3"}],"recommendation":"reject","confidential_remarks":"The inductive-generation gap is not a presentation issue: the formal model in §3.2 has a fixed categorical support, and §3.4 lacks any node-arrival mechanism. Adding such a mechanism would require a substantive change to the model, so I cannot recommend major revision as a path within scope. The exponential inter-event time assumption and the uncontrolled link-prediction baseline comparison are additional concerns that would need to be addressed in any future submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one genuinely new thing here is the direct factorization of a temporal interaction as p(src)p(dst|src)p(t,e|src,dst), with a deep probabilistic decoder over TGN embeddings. That is a real departure from random-walk or motif-rate generators, and it lets the model handle edge features of arbitrary size, which no prior CTDG generator did. The paper is also clearly written, the autoregressive sampling scheme is sensible, and the link-prediction result—DG-Gen beating TGN while using TGN as its encoder—is a genuinely interesting signal that the decoder adds something.\n\nThe soft spots are real, though. The biggest one is the inductive claim. The model defines p(src) and p(dst|src) as Categorical distributions over a fixed node set, and generation starts from an empty graph with no node-arrival mechanism. Nothing in the paper explains how an unseen node ID gets into that support. In practice, the node vocabulary is whatever was loaded at training time, so generating a graph with nodes outside the training set is not supported by the described math. If the authors want to keep the inductive claim, they need an explicit node-generation step and experiments that actually report whether any generated node ID is absent from the training set. Without that, the honest claim is transductive generation on a fixed node set, which is still useful but is not what the abstract promises.\n\nSecond, the exponential distribution for inter-event times is an assumption slipped in as if it were minor. The paper says the method is \"largely assumption-free,\" but exponential inter-event times are a strong parametric choice and it is never tested against the data. That should be checked, or the distribution should be made more flexible.\n\nThird, the evaluation has some sloppiness. Table 2 reports median absolute errors without error bars or significance, and several link-prediction baselines are imported from earlier work rather than rerun under a common protocol. The feature-distribution results in Table 3 do have standard deviations, which helps, but the topology metrics need the same treatment.\n\nOverall, I think the core idea is novel and the paper is a solid first step. The inductive claim is load-bearing for the stated contributions, so if it stays, the paper needs real revision. But this deserves referee time, not a desk reject. For someone working on temporal graph generation, it is worth reading as a new angle, even if they should not cite the inductive claim until it is actually demonstrated.","headline":"A novel direct factorization for CTDG generation, but the inductive-generation claim is not supported by the described mechanism; worth a serious revision, not a desk reject.","tokens_in":18140,"tokens_out":2522,"would_cite":false,"duration_ms":25213,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The central claim is that temporal graphs can be generated event-by-event by sampling a factored joint probability over source, destination, time, and edge features, with no static-graph or snapshot representation.","keywords":["continuous-time dynamic graphs","graph generation","temporal interaction modeling","probabilistic factorization","autoregressive generation","temporal graph networks","link prediction","edge feature generation"],"falsifier":"After training on a graph with node set $V_{\\mathrm{train}}$, initialize the encoder with a held-out node's raw features or no memory, and ask the model for $p(\\mathrm{src})$ and $p(\\mathrm{dst}\\mid \\mathrm{src})$ involving that node; if the architecture cannot issue such a query, or if a generated graph of the same length as the test partition never contains any node outside $V_{\\mathrm{train}}$, then the inductive-node-generation claim is not realized.","tokens_in":17150,"feed_emoji":"🕸️","tokens_out":8893,"duration_ms":69881,"temperature":0.7,"pith_summary":"The paper introduces DG-Gen, a generative framework for continuous-time dynamic graphs that treats each temporal interaction as a joint probability over its four attributes: source node, destination node, timestamp, and edge-feature vector. The joint probability is factored as $p(\\mathrm{src}, \\mathrm{dst}, t, e) = p(\\mathrm{src})\\,p(\\mathrm{dst}\\mid \\mathrm{src})\\,p(t,e\\mid \\mathrm{dst},\\mathrm{src})$, and a deep probabilistic decoder estimates the parameters of each factor from temporal node embeddings. The central claim is that sampling these factors autoregressively produces synthetic dynamic graphs that are statistically similar to the source graph without duplicating its edges, and that the same learned probabilities can be read directly as a link-prediction model. On five datasets the authors report that DG-Gen beats the only prior inductive CTDG generator on graph-fidelity metrics and matches or beats dedicated link-prediction models on four of the five datasets.","feed_headline":"One probability model generates dynamic graphs and predicts links","feed_subtitle":"Factored event probabilities let DG-Gen write synthetic temporal graphs that match source statistics.","key_machinery":"The central object is the factored interaction probability $p(\\mathrm{src}, \\mathrm{dst}, t, e) = p(\\mathrm{src})\\,p(\\mathrm{dst}\\mid \\mathrm{src})\\,p(t,e\\mid \\mathrm{dst},\\mathrm{src})$, which decomposes a temporal edge into source selection, source-conditioned destination selection, and event content (time and edge features). Each factor is parameterized by a deep decoder over temporal node embeddings: the Reshape module scores a node as a source, the Product module scores a source-destination pair, the Merge module combines two node embeddings into a vector $h_0$, and the Time+MSG module, initialized with $h_0$, outputs the parameters of the inter-event-time distribution and the edge-feature distributions. The decoder is trained by minimizing the negative log-likelihood of observed interactions. At inference, sampling from the three factors in sequence and updating node memories after each batch produces the synthetic graph.","core_discovery":"The discovery is that a continuous-time dynamic graph can be generated event-by-event without ever constructing a static graph or discrete snapshots. DG-Gen writes the probability of a temporal edge as $p(\\mathrm{src}, \\mathrm{dst}, t, e) = p(\\mathrm{src})\\,p(\\mathrm{dst}\\mid \\mathrm{src})\\,p(t,e\\mid \\mathrm{dst},\\mathrm{src})$, where $p(\\mathrm{src})$ is a Categorical distribution over possible source nodes, $p(\\mathrm{dst}\\mid \\mathrm{src})$ is a Categorical distribution over destinations conditioned on the source, and $p(t,e\\mid \\mathrm{dst},\\mathrm{src})$ combines an Exponential inter-event-time model with Categorical or Gaussian-mixture models for edge features. A temporal encoder supplies node memories and embeddings, and the decoder's modules map those embeddings into the distribution parameters. Inference starts from an empty graph and samples sources, then destinations, then event times and features, updating memories after each generated batch. The same factorization yields link prediction for free as $p(\\mathrm{dst}\\mid \\mathrm{src})$, and the reported experiments show it matching or surpassing dedicated discriminative baselines while also generating edge features of arbitrary length.","pith_inferences":["An implicit consequence the authors do not develop is that the factorization separates generation into three channels: who acts, whom they act on, and what the event contains; each factor could be reweighted or conditioned independently, for example to simulate counterfactual graphs with the same topology but different timing or feature distributions.","A concrete extension would pair the decoder with a different temporal encoder, since the decoder consumes only embeddings; the framework predicts that encoder quality transfers directly to generation fidelity, which could be tested by swapping in another encoder.","The inductive claim is only as strong as the node set: because $p(\\mathrm{src})$ is defined over nodes with learned embeddings and generation starts with empty memory, the paper does not specify how a truly new node enters the set, so extending the method to genuinely unseen nodes would require a node-arrival model or an embedding-initialization rule."],"forward_implications":["If the central claim is correct, synthetic continuous-time graphs can be generated event-by-event without constructing an adjacency matrix, so generation can scale to graphs with hundreds of thousands or millions of timestamps.","The same trained model performs link prediction without fine-tuning, since $p(\\mathrm{dst}\\mid \\mathrm{src})$ is exactly the conditional probability that a given source connects to a given destination at the current temporal state.","Edge features of arbitrary length, both categorical and numerical, can be generated from learned conditional distributions, a capability the paper says previous CTDG generators lacked.","Generated graphs are original in the edge-overlap sense used by the baseline literature, so they can support data augmentation and obfuscation without re-releasing source edges.","On the five datasets studied, the fidelity results place DG-Gen ahead of the only prior inductive CTDG baseline on most topology metrics, with the largest gains in closeness centrality."],"supporting_citations":[{"why":"Supplies the temporal encoder used to compute node memories and embeddings from raw interaction data.","marker":"Rossi et al. 2020"},{"why":"Provides TIGGER/TIGGER-I, the baseline inductive CTDG generator, and the edge-overlap originality measure used for comparison.","marker":"Gupta et al. 2022"},{"why":"Defines TagGen, the static-graph-plus-temporal-edge approach whose limitations motivate direct event modeling.","marker":"Zhou et al. 2020"},{"why":"Defines Dymond, the motif-based snapshot generative baseline that DG-Gen contrasts with on topological assumptions and scalability.","marker":"Zeno, La Fond, and Neville 2021"},{"why":"Defines the inductive sampling protocol used to evaluate link prediction on unseen edges.","marker":"Poursafaei et al. 2022"},{"why":"Supplies JODIE baseline and the Reddit, Wikipedia, MOOC, and LastFM datasets used in the experiments.","marker":"Kumar, Zhang, and Leskovec 2019"},{"why":"Grounds the notion of inductive learning that the paper uses to claim generation of unseen nodes.","marker":"Hamilton, Ying, and Leskovec 2017"}],"fun_headline_variants":["Event-by-event dynamic graph generation without snapshots","One probability model writes temporal graphs and predicts links","Joint probabilities generate dynamic graphs and improve link prediction","Sampling events generates synthetic dynamic graphs directly","Directly model edge probabilities to generate continuous-time graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's inductive-generation claim rests on the assumption that a node not present during training can receive a meaningful embedding and probability under $p(\\mathrm{src})$ and $p(\\mathrm{dst}\\mid \\mathrm{src})$, but no node-arrival or embedding-initialization mechanism for unseen nodes is specified.","fun_headline_variants_meta":{"raw":{"variants":["Event-by-event dynamic graph generation without snapshots","One probability model writes temporal graphs and predicts links","Joint probabilities generate dynamic graphs and improve link prediction","Sampling events generates synthetic dynamic graphs directly","Directly model edge probabilities to generate continuous-time graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1338,"prompt_tokens":957,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":310}},"tokens_in":573,"tokens_out":381,"duration_ms":3520,"temperature":1.0,"reasoning_tokens":310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:17:43.192278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"After training on a graph with node set $V_{\\mathrm{train}}$, initialize the encoder with a held-out node's raw features or no memory, and ask the model for $p(\\mathrm{src})$ and $p(\\mathrm{dst}\\mid \\mathrm{src})$ involving that node; if the architecture cannot issue such a query, or if a generated graph of the same length as the test partition never contains any node outside $V_{\\mathrm{train}}$, then the inductive-node-generation claim is not realized.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides TIGGER/TIGGER-I, the baseline inductive CTDG generator, and the edge-overlap originality measure used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines TagGen, the static-graph-plus-temporal-edge approach whose limitations motivate direct event modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines Dymond, the motif-based snapshot generative baseline that DG-Gen contrasts with on topological assumptions and scalability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the inductive sampling protocol used to evaluate link prediction on unseen edges."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies JODIE baseline and the Reddit, Wikipedia, MOOC, and LastFM datasets used in the experiments."}],"review_version":1}