{"id":"b6efa9df-c8a2-42ab-817b-b252f5309fab","arxiv_id":"1908.01623","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A graph-biased temporal point process reports better propagation prediction than RMTPP, but the evaluation is compromised by graph embeddings computed from the full dataset before cross-validation.","lead":"This paper adds graph structure to a neural point process model for predicting how events spread across social networks and websites. The reported gains are not trustworthy because the graph features are built from the same sequences used for testing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Graph embeddings are fitted to the full dataset before the 10-fold split, so the graph bias term can memorize test-fold transitions; the reported GBTPP-vs-RMTPP gains are not trustworthy.","rationale":"I read the paper as proposing an extension of RMTPP that uses graph structure both as a bias term and as side information. For the central claim to hold, the evaluation must compare GBTPP and RMTPP under conditions where GBTPP does not see test-fold outcomes through the graph representation. The paper's own protocol contradicts this: Section III-A says the graph is learned from observed propagation sequences, and Section IV-B estimates the Meme adjacency matrix from all meme cascades before the 10-fold split. Because the bias term in Eq. 7 contains p(y_n,y_k) derived from that adjacency, test-fold transition counts can directly inflate GBTPP's predictions. This is not a disagreement with the model's design; it is an internal flaw in the experimental support. The Eq. 9 normalization error is an additional correctness concern that independently undermines the time-RMSE comparison. The synthetic data setup and the NRPP ablation are useful elements, but they do not repair the real-data evaluation as reported. Since my reading supports the reader's REJECT without further adjustment, I mark the verdict as UNCHANGED.","tokens_in":17212,"tokens_out":5150,"duration_ms":49370,"concrete_test":"Re-run the MemeTracker and Higgs evaluations with the graph-construction step nested inside cross-validation: for each of the 10 folds, compute A and the node embeddings from the training sequences only, then evaluate on the held-out fold (or use an independent graph source not derived from the target cascades). Compare GBTPP accuracy and time RMSE with RMTPP on identical splits. If the advantage shrinks to within one standard deviation or reverses, the reported improvement is an artifact of test-fold leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-A, the adjacency matrix A_{ij}=N_{ij}/N_{max} is computed from all observed propagation sequences, and node embeddings are trained to reconstruct A via Eq. 2. Section IV-B then describes 10-fold cross-validation by splitting the sequence collection, with no statement that A or the embeddings are re-estimated inside each fold. Since Eqs. 6-7 insert p(y_n,y_k) into the next-node probability, the graph bias term is effectively a lookup over transition counts that include the test folds. A test-fold transition that has occurred before receives an inflated score, and this advantage is not available to RMTPP or NRPP. The same contamination affects the top-K curve in Fig. 6. The central empirical claim that GBTPP 'achieves better performance than RMTPP, especially on real-world datasets' therefore rests on an unfair comparison as reported. A second, independent correctness issue is in Eq. 9: after integrating the intensity of Eq. 8, the lower-limit term should be (1/w) exp(vh^T h_{n-1} + vy^T y_n + b_t); as written with vt^T h_{n-1}+b_t, the density is not correctly normalized. This affects the time-likelihood component of Eq. 12 and the RMSE results. The model idea is plausible and the NRPP ablation is a reasonable control, but the central empirical claim is not supported by the experiments as described.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Graph Biased Temporal Point Process (GBTPP) for modeling event propagation in a graph. The method first learns node embeddings by reconstructing an adjacency matrix A_{ij}=N_{ij}/N_{max} estimated from observed propagation sequences, then feeds these embeddings into a recurrent marked temporal point process. The next-node probability adds a graph bias term p(y_n,y_k) modulated by a history-dependent scale, and the time intensity adds a direct-influence term v_y^T y_n. The authors evaluate on a synthetic Hawkes dataset, the Higgs Twitter dataset, and MemeTracker, comparing against RMTPP, Markov chains, point-process baselines, and an NRPP ablation, and report improved accuracy and RMSE.","tokens_in":17542,"tokens_out":7318,"duration_ms":67067,"significance":"If the reported evaluation were sound, the paper would make a useful contribution by separating direct graph influence from indirect history influence in neural temporal point processes, and the NRPP ablation is a sensible control for isolating the graph bias term. The experimental setup covers a synthetic and two real datasets, which is appropriate for the claim. However, the graph representation step appears to use test-fold information, and Eq. (9) is internally inconsistent, so the empirical claims are not currently supported. The underlying modeling idea remains plausible and could be of interest after a corrected within-fold evaluation.","major_comments":[{"comment":"The graph representation is learned from the full dataset before the 10-fold split. In Section III-A, A_{ij}=N_{ij}/N_{max} is computed from all observed propagation sequences C, and the node embeddings y_i are fitted to A via Eq. (2). Section IV-B then partitions C into folds without stating that A or the embeddings are re-estimated inside each training fold; for MemeTracker this is explicit ('Firstly, the adjacent matrix is estimated by A_{ij}=N_{ij}/N_{max}... Given adjacent matrix A, the graph embedding {y_k} is learned... then the GBTPP model is trained'), and the Higgs description follows the same order. Because Eq. (6) inserts p(y_n,y_k) into the next-node probability, the graph bias term can act as a lookup over transition counts that include test-fold events. This gives GBTPP an information advantage that RMTPP and NRPP do not have and invalidates the comparisons in Table I and Fig. 6. Please re-estimate A and the node embeddings separately inside each training fold, or use a graph source that is independent of the propagation labels, and report the results.","section":"Section III-A / IV-B"},{"comment":"Equation (9) is not the correct density for the intensity in Eq. (8). Integrating over [t_{n-1}, t] gives a lower-limit term (1/w) exp(v_h^T h_{n-1} + v_y^T y_n + b_t), but Eq. (9) writes (1/w) exp(v_t^T h_{n-1} + b_t), omitting the direct-influence term v_y^T y_n; the first exponent also uses v_t^T where Eq. (8) uses v_h^T. As written the density is not normalized, so the time log-likelihood in Eq. (12) and the reported time RMSE are not trustworthy. Please correct the expression, harmonize the notation, and re-run the time-prediction experiments.","section":"Section III-B, Eq. (9)"}],"minor_comments":[{"comment":"The experimental section does not report the embedding dimension d, hidden state dimension H, learning rate, BPTT length b, or number of training epochs; these details are needed to reproduce the results.","section":"Section IV-B"},{"comment":"No error bars are shown in the top-K curves of Fig. 6, and no significance tests accompany Table I; given the small differences on the synthetic accuracy (e.g., 46.82 vs 47.26 with overlapping standard deviations), 'better performance' should be qualified.","section":"Section IV-C"},{"comment":"Eq. (2) is described as KL-divergence minimization, but the expression is a weighted cross-entropy; the relationship to the empirical distribution \\hat{p}^{(1)} should be clarified.","section":"Section III-A"},{"comment":"There are numerous typos, such as 'defining the the grain' in Section II-A, and inconsistent use of v_t and v_h in Section III-B; a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The leakage in the graph representation step is the central issue; if the authors cannot demonstrate that the gains survive a within-fold estimation of A and the embeddings, the paper should not be accepted. The Eq. (9) normalization error is independent and must also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the model idea is a reasonable, incremental extension of RMTPP, but the evaluation as written does not support the headline claim. The graph embeddings are learned from an adjacency matrix built from all propagation sequences before the data split, so the graph bias term can peek at test-fold transitions.\n\nWhat is genuinely new: the two changes to RMTPP are real. The graph bias term in the softmax separates direct influence between connected nodes from indirect history influence, and the node embeddings are also fed into the history encoder as side information. That combination is not in the cited baselines. The NRPP ablation is a decent control, and the top-K curves are a useful diagnostic. I do not see evidence of intentional misrepresentation; the paper describes its split procedure clearly. The problem is the protocol.\n\nSoft spots, in proportion: the leakage is the load-bearing one. Section III-A defines A_ij = N_ij / N_max from all observed propagation sequences, and the node embeddings are fitted to that adjacency matrix. Section IV-B then splits the sequence collection for 10-fold cross-validation, with no statement that A or the embeddings are re-estimated inside each fold. Equations 6 and 7 insert p(y_n, y_k) into the next-node probability, so a test-fold transition that already appeared in the full adjacency matrix receives an inflated score. RMTPP and NRPP do not get that advantage. This contaminates both the accuracy numbers and the top-K curves, and it is not a minor procedural omission—it is the load-bearing part of the empirical claim.\n\nThere is also an independent correctness issue in Eq. 9. After integrating the intensity in Eq. 8, the lower-limit term should include vy^T y_n inside the exponential. As printed, the last term has only vh^T h_{n-1} + b_t, so the density is not correctly normalized. This affects the time-likelihood component of Eq. 12 and hence the RMSE results. It may be a typo, but it needs to be fixed and the experiments rerun.\n\nBottom line: this paper is for someone working on neural temporal point processes and cascade prediction who wants the formulation and a cautionary example. I would not desk reject it if a journal asked me—send it for review with a clear request to rerun the experiments with per-fold graph estimation and to correct Eq. 9—but I would not trust the reported numbers as they stand.","headline":"The graph-bias idea is a plausible incremental extension of RMTPP, but the evaluation leaks test-fold transitions into the graph embeddings, so the reported gains are not trustworthy as written.","tokens_in":18057,"tokens_out":3207,"would_cite":false,"duration_ms":35147,"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 paper claims that separating direct node influence from historical influence in a temporal point process improves propagation prediction on real-world networks.","keywords":["temporal point process","event propagation","graph representation learning","node embedding","recurrent neural network","propagation prediction","social network analysis","meme tracking"],"falsifier":"Re-estimate the adjacency matrix $A_{i,j}$ and node embeddings using only training sequences in each 10-fold split, then retrain GBTPP and compare node accuracy and time RMSE against RMTPP. If the advantage shrinks or reverses, the reported gains came from test-fold information entering the graph bias term.","tokens_in":17030,"feed_emoji":"📢","tokens_out":3848,"duration_ms":38139,"temperature":0.7,"pith_summary":"The paper proposes a graph-biased temporal point process (GBTPP) for modeling how events propagate through a network, such as retweets on Twitter or memes between websites. It separates the direct influence of the currently reached node, captured by a graph bias term built from node embeddings, from the indirect influence of the entire propagation history, modeled by a recurrent neural network. Using a synthetic Hawkes dataset, the Higgs Twitter dataset, and the MemeTracker dataset, the paper claims GBTPP improves both next-node prediction accuracy and propagation-time RMSE over the recurrent marked temporal point process (RMTPP) baseline, with larger gains on real-world data.","feed_headline":"Graph-aware model beats RMTPP on propagation prediction","feed_subtitle":"GBTPP adds node-proximity bias and graph embeddings, improving next-node accuracy and timing on Twitter and MemeTracker.","key_machinery":"The central object is the graph bias term $b(h_{n-1}, y_n, y_k) = \\mathrm{ReLU}(U^h_{n,:}h_{n-1})\\, p(y_n, y_k)$, where $p(y_n, y_k) = \\sigma(y_n^{s\\top} y_k^e)$ is the first-order proximity learned from node embeddings. The embeddings are obtained by minimizing the KL divergence between this embedding-based proximity and an empirical adjacency matrix $A_{i,j} = N_{ij}/N_{\\max}$ built from normalized propagation counts, as in Eq.\\ (2). The ReLU factor lets the history modulate how strongly the direct edge influences the next node, and the node embedding vectors are also injected into the recurrent history embedding of Eq.\\ (3) as side information.","core_discovery":"GBTPP predicts the next propagation node $v_{n+1}$ and time $t_n$ given history $H_{n-1}$ and current node $v_n$. The probability of each candidate node $k$ is a softmax over $V^h_{k,:}h_{n-1} + b(h_{n-1}, y_n, y_k) + b^h_k$, where the graph bias term $b(h_{n-1}, y_n, y_k) = \\mathrm{ReLU}(U^h_{n,:}h_{n-1})\\, p(y_n, y_k)$ multiplies a history-dependent scale factor by the first-order proximity between the current node and the candidate. The conditional intensity in Eq.\\ (8) likewise adds an explicit direct-influence term $v_y^\\top y_n$ to the history term. The full model is trained by maximizing the joint log-likelihood of propagation node and time via backpropagation through time, and the paper reports that this explicit decomposition of direct and indirect influence yields better accuracy and lower RMSE than RMTPP, especially on Higgs and MemeTracker data.","pith_inferences":["A critical control study would re-estimate the adjacency matrix and node embeddings using only training-fold sequences in each 10-fold split; the reported advantage over RMTPP may shrink if the graph bias currently encodes test-fold transition counts.","The same bias mechanism could be combined with continuous-time graph neural networks or temporal graph embeddings to handle evolving network structure, going beyond the static-graph setting.","The exponential kernel assumption in the intensity function is a parametric choice; replacing it with a learned kernel might further improve time prediction on irregular propagation patterns.","The first-order proximity bias only uses direct edges; extending the bias to second-order proximity could capture the paper's motivating example of $V_2 \\to V_4 \\to V_5$ without relying solely on the RNN history."],"forward_implications":["If the claim holds, propagation modeling in social and information networks can be improved by explicitly encoding graph structure into neural temporal point processes.","The direct-versus-indirect decomposition offers a modular design: any RNN-based point process can be augmented with a pre-learned graph bias without changing its recurrent core.","The history-dependent scale factor $\\mathrm{ReLU}(U^h_{n,:}h_{n-1})$ effectively gives each static edge a time-varying strength, which may help in cascades where influence wanes or surges.","The top-$K$ prediction formulation positions the model as a recommender that can suggest likely next nodes and their propagation times.","The framework extends naturally to dynamic graphs and to event sequences with richer semantic side information, as the paper itself points to as future work."],"supporting_citations":[{"why":"Supplies the temporal point process framework that GBTPP extends.","marker":"[4]"},{"why":"Defines the recurrent marked temporal point process (RMTPP) baseline whose history embedding and intensity form GBTPP modifies.","marker":"[6]"},{"why":"Provides the Higgs Twitter retweet dataset used in the real-world evaluation.","marker":"[5]"},{"why":"Provides the MemeTracker meme-propagation dataset used in the evaluation.","marker":"[17]"},{"why":"Supplies the multidimensional Hawkes generative model used to create the synthetic dataset.","marker":"[44]"}],"fun_headline_variants":["GBTPP adds graph bias to point processes for sharper propagation forecasts","Direct influence via graph proximity boosts temporal point process accuracy","Graph-biased point process improves next-node and timing prediction","GBTPP: graph embeddings and proximity bias refine event propagation","Explicit direct influence from graph structure sharpens propagation models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The graph representation is learned from the complete observed dataset, including test-fold transition counts, before the cross-validation splits are made; the central performance claim collapses if the adjacency matrix must be estimated from training folds only.","fun_headline_variants_meta":{"raw":{"variants":["GBTPP adds graph bias to point processes for sharper propagation forecasts","Direct influence via graph proximity boosts temporal point process accuracy","Graph-biased point process improves next-node and timing prediction","GBTPP: graph embeddings and proximity bias refine event propagation","Explicit direct influence from graph structure sharpens propagation models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000465,"raw_usage":{"total_tokens":2307,"prompt_tokens":916,"completion_tokens":1391,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":1308}},"tokens_in":532,"tokens_out":1391,"duration_ms":10382,"temperature":1.0,"reasoning_tokens":1308,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:07:15.014389+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate the adjacency matrix $A_{i,j}$ and node embeddings using only training sequences in each 10-fold split, then retrain GBTPP and compare node accuracy and time RMSE against RMTPP. If the advantage shrinks or reverses, the reported gains came from test-fold information entering the graph bias term.","supporting_citations":[{"cited_title":"Daley and D","cited_arxiv_id":null,"evidence_quote":"Supplies the temporal point process framework that GBTPP extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the recurrent marked temporal point process (RMTPP) baseline whose history embedding and intensity form GBTPP modifies."},{"cited_title":"De Domenico, A","cited_arxiv_id":null,"evidence_quote":"Provides the Higgs Twitter retweet dataset used in the real-world evaluation."},{"cited_title":"Leskovec, L","cited_arxiv_id":null,"evidence_quote":"Provides the MemeTracker meme-propagation dataset used in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multidimensional Hawkes generative model used to create the synthetic dataset."}],"review_version":1}