{"id":"4fad077b-56c9-4ca2-ad2a-903032a49f3e","arxiv_id":"2505.15746","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"HTGN adds hyperedge memory and hypergraph convolution to temporal GNNs, claiming better dynamic link prediction and lower memory cost, but the reported results are undermined by data inconsistencies and invalid proofs.","lead":"HTGN is a temporal graph neural network that builds hyperedges from groups of recent interactions, stores them in a memory module, and uses hypergraph convolution for link prediction. The paper claims better accuracy and up to 50% lower GPU memory, but the experimental tables contain impossible duplicate rows and the theoretical proofs are unsound.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 copies GraphMixer, TGAT, and DyRep baseline values from Table 1 onto different datasets; the claimed link-prediction gains over baselines are not supported.","rationale":"I read the paper as making two coupled claims: higher-order structure improves dynamic link prediction, and hyperedge memory cuts GPU cost. The prediction claim is supported by Tables 1 and 2; the memory claim by Figure 3. The load-bearing condition is that the baseline numbers in those tables are real measurements. They are not: Table 2 reuses Table 1's GraphMixer, TGAT, and DyRep values on datasets with no overlap. This is not an interpretive disagreement; it is a factual inconsistency. The reader's weakest_assumption focused on the validity of the hyperedge-construction theorem, and the reader's rationale also flagged these table errors. I agree the theorem proof is problematic, but the duplicated baseline rows are the more directly falsifiable defect in the central empirical claim. A single rerun of the duplicated baseline rows would settle the issue. I therefore keep the reader's REJECT verdict; no adjustment is needed.","tokens_in":15443,"tokens_out":7629,"duration_ms":64226,"concrete_test":"Request the original experimental logs or code and rerun GraphMixer, TGAT, and DyRep on tgbl-wiki, Reddit, and LastFM under the protocol in §6.1. If the resulting MRR values do not match Table 2, which currently repeats Table 1's UCI, Enron, and SocialEvo entries, then the heterogeneous baseline comparison is invalid and HTGN's rank and average improvement must be recomputed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim — that HTGN achieves superior dynamic link prediction performance — rests on Tables 1 and 2. Table 2 is unreliable: for the heterogeneous datasets tgbl-wiki, Reddit, and LastFM, the GraphMixer row is 17.6/20.4, 21.9/23.7, 35.5/35.3 in Val/Test, identical to Table 1's GraphMixer values for UCI, Enron, and SocialEvo. The same duplication occurs for TGAT (20.4/23.5, 36.3/35.8, 40.4/41.6) and DyRep (7.2/5.0, 11.2/12.7, 15.6/19.4). These are disjoint datasets; exact equality to one decimal across three datasets cannot arise by chance. Consequently, HTGN's reported rank of 1.3 on heterogeneous graphs and the claimed 5% average improvement over the second-best method are computed against baseline numbers that are not real measurements for those datasets. Table 3 compounds the integrity problem: the tgbl-comment row reports 125,235 edges but 30,146,596 unique edges, an internal contradiction. Without corrected tables, or code and data, the advertised performance advantage over pairwise baselines is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HTGN, a temporal graph neural network that constructs hyperedges from temporal event streams (maximal cliques in short snapshots for homogeneous graphs, bipartite projections with a time threshold for heterogeneous graphs), stores hyperedge features in a dynamic memory module, and applies hypergraph convolution to compute node embeddings for dynamic link prediction. The authors claim that HTGN achieves superior link prediction performance on ten real-world datasets while reducing GPU memory by up to 50%, and they provide two theoretical results: Theorem 1, which asserts that hyperedge-construction accuracy decreases with snapshot duration, and Theorem 2, which asserts that HTGN is strictly more expressive than pairwise message-passing temporal GNNs.","tokens_in":15764,"tokens_out":5684,"duration_ms":43957,"significance":"If the claims were substantiated, the work would be a meaningful contribution to temporal graph learning: it would be the first temporal GNN to demonstrate that exploiting higher-order structures can simultaneously improve accuracy and reduce memory cost. The idea of constructing hyperedges dynamically from event streams and allocating memory units to hyperedges rather than to individual nodes is interesting and practically motivated, and the paper makes its algorithms explicit. However, the empirical evidence is compromised by apparent duplication of baseline numbers across different datasets and by impossible dataset statistics, and the theoretical proofs contain load-bearing gaps. As presented, the central claims are not established, so the current significance is mostly in the proposal's potential rather than in its validated results.","major_comments":[{"comment":"The baseline rows for GraphMixer, TGAT, and DyRep in Table 2 reproduce exactly, to one decimal place, the values reported in Table 1 for UCI, Enron, and SocialEvo, even though the datasets in Table 2 (tgbl-wiki, Reddit, LastFM) are different. For example, GraphMixer is 17.6/20.4, 21.9/23.7, 35.5/35.3 in both tables, and TGAT is 20.4/23.5, 36.3/35.8, 40.4/41.6 in both tables. Such exact equality across disjoint datasets cannot arise from measurement, so the baseline results for heterogeneous graphs are not credible. Since HTGN's claimed rank of 1.3 and the average improvement of 5% are computed against these numbers, the central empirical claim of superior link prediction performance is unsupported.","section":"Section 6.2, Tables 1 and 2"},{"comment":"The dataset statistics in Table 3 contain impossible entries. The tgbl-comment row reports # Edges = 125,235 and # Unique Edges = 30,146,596, which is impossible because unique edges cannot exceed total edges. The tgbl-flight row reports 67,169,570 edges and 18,143 nodes but only 3,125 unique edges, which is implausible for a network of this size and is identical to Enron's unique edge count, suggesting a copy-and-paste error. These inconsistencies indicate that the experimental data pipeline is unreliable and need to be corrected before any performance claims can be evaluated.","section":"Table 3"},{"comment":"The proof of Theorem 1 misapplies Lemma A.1 (the PAC-Bayes inequality from Germain et al.). The lemma states E_{h~Q}[ψ(h)] ≤ DKL(Q||P) + ln E_{h~P}[e^{ψ(h)}], which is an upper bound consisting of the KL divergence plus a log-partition term. In line (16), the proof instead claims E_Q[2^{-m}] ≤ DKL(Q||P) − E_P[m ln(e/2)], subtracting a positive quantity (since ln(e/2) > 0). This reverse direction is not justified by the lemma and invalidates the derivation. Consequently, Theorem 1 does not establish that shorter snapshots yield higher hyperedge-construction accuracy, so the heuristic in Algorithm 1 lacks the claimed theoretical support.","section":"Appendix A, proof of Theorem 1"},{"comment":"The proof of Theorem 2 is not rigorous. It claims that HTGN uses an 'injective aggregation function' to combine hyperedge features, but Eq. (3) defines the aggregation as a sum of MLP outputs weighted by an exponential decay, which is not injective in general. The example in Figure 5 is underspecified: the node features and timestamps are not given in a way that verifies the claim that MP-TGN cannot distinguish the two graphs, and the argument that every link used by MP-TGN is also used by HTGN assumes a specific message-passing order that is not formalized. Thus the strict-expressiveness result is not established.","section":"Appendix B, proof of Theorem 2"}],"minor_comments":[{"comment":"Equation (3) contains a typo: 'MLP(m[E]||)' has a dangling concatenation symbol and no second argument; it should likely be 'MLP(m[E])' or 'MLP(m[E] ∥ ...)' with the intended term specified.","section":"Section 5.1, Eq. (3)"},{"comment":"In Definition 4.3, the union symbol is missing in 'SK a=1 Ca = V'; it should read '∪_{a=1}^K C_a = V'.","section":"Definition 4.3"},{"comment":"The citation for negative sampling, Patania et al. (2017), appears unrelated to negative sampling in temporal link prediction; this reference should be replaced with the appropriate source.","section":"Section 6.1"},{"comment":"In Algorithm 2, line 7, 'Find the node in |E(v)|' should read 'Find the node in E[v]' since E(v) is a set, not a numeric quantity.","section":"Appendix H, Algorithm 2"},{"comment":"The text refers to 'msgg()' while Eq. (5) defines 'msgd()'; the notation should be made consistent.","section":"Section 5.1, Eqs. (4)-(7)"},{"comment":"The y-axis label 'MRR change' is ambiguous: it is unclear whether the numbers represent percentage-point differences or relative improvements, and the caption should state the normalization.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The duplication of baseline values across Tables 1 and 2 is a serious data-integrity concern that goes beyond ordinary experimental error. I would recommend that the editor request the raw experimental logs and code before considering any revised version. The paper also lacks a reproducibility statement or link to code, which compounds the problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: HTGN is a plausible idea—build hyperedges from short-window snapshots, store hyperedge memory, and run hypergraph convolution—but the paper's own evidence undercuts it. The key tables are internally inconsistent, and the theoretical proofs don't work. I wouldn't trust the reported results.\n\nWhat's genuinely new: the combination of temporal memory with hypergraph convolution to capture group interactions and reduce memory is not something I've seen in the TGNN literature. The hyperedge construction algorithm, while heuristic, is a reasonable way to get higher-order structure without enumerating all cliques globally. If it worked, it would be a useful practical contribution.\n\nThe problems are hard to ignore. Table 2's GraphMixer, TGAT, and DyRep rows are identical to Table 1's values for different datasets. That's not a typo; it's a direct copy-paste of baseline numbers onto datasets they weren't run on. The same happens for TCL and others. The average ranks and '5% improvement' claims are computed on top of these fabricated numbers. Table 3 is equally bad: tgbl-comment has 125,235 edges but 30,146,596 unique edges, which is impossible. The theorem proofs are also suspect. Appendix A's proof of Theorem 1 applies the PAC-Bayes bound in the wrong direction—Lemma A.1 gives E_Q[...] ≤ D_KL + ln E_P[...], but the proof introduces a minus sign. The statement of the theorem itself is odd, with an upper bound that could be negative. Theorem 2 is a sketch: it relies on the assumption that the constructed hyperedges correspond to real higher-order interactions, and the example in Figure 5 is not fully analyzed.\n\nThere's no code or data, so there's no way to check any of this independently.\n\nI don't think this is a 'sound idea with minor flaws.' The flaws are load-bearing: the accuracy claims rest on broken tables, the memory claims are only shown in a log-scale figure with no numbers, and the theory is not proven. A serious editor would desk reject, or at minimum require a major overhaul with corrected tables, code, and a proper proof.\n\nFor a reader: if you're interested in higher-order temporal graph learning, the idea is worth a thought, but cite it with caution. I would not bring it to reading group as a model paper; maybe as a cautionary example.\n\nRecommendation: do not send to peer review in its current form. If the authors can fix the tables and provide code with reproducible numbers, it might become a real submission.","headline":"Interesting architecture, but the experimental tables are unreliable and the theory doesn't hold; not ready for review.","tokens_in":16288,"tokens_out":4932,"would_cite":false,"duration_ms":41531,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By treating short-lived group interactions as hyperedges, HTGN claims both higher dynamic link prediction accuracy and up to 50% lower memory use than pairwise temporal GNNs.","keywords":["temporal graph neural networks","dynamic link prediction","higher-order structures","hypergraph construction","maximal cliques","memory efficiency","expressiveness","temporal graphs"],"falsifier":"Train HTGN and a strong pairwise baseline such as TGN or TNCN on a temporal network generated with no group interactions, using equal-sized embeddings; if HTGN still matches or beats the baseline, its edge is not coming from higher-order structure. Alternatively, run HTGN with snapshot sizes large enough that Theorem 1 predicts degraded hyperedge accuracy; if link prediction MRR does not drop, the hyperedge construction step is not driving the result.","tokens_in":15244,"feed_emoji":"🕸️","tokens_out":6280,"duration_ms":49422,"temperature":0.7,"pith_summary":"The paper sets out to show that temporal graph neural networks can be both more accurate at dynamic link prediction and cheaper to run if they represent group interactions as hyperedges instead of expanding them into many pairwise edges. It introduces HTGN, which builds hyperedges from maximal cliques in short snapshots of the most recent edges, stores a single memory feature per hyperedge, and computes node embeddings by hypergraph convolution. The paper reports that HTGN ranks first or second on 9 of 10 real-world datasets and reduces GPU memory by 30–50 percent compared with memory-based pairwise baselines. It also proves that HTGN is strictly more expressive than pairwise message-passing temporal GNNs, giving a theoretical reason for the performance gain. The significance, if the claim holds, is that expressiveness and efficiency need not trade off in dynamic link prediction.","feed_headline":"Group-aware temporal GNN beats pairwise rivals, cuts memory 50%","feed_subtitle":"Storing group interactions as hyperedges boosts dynamic-link prediction on 9 of 10 real-world networks.","key_machinery":"The load-bearing mechanism is Algorithm 1, which collects incoming temporal edges into a snapshot and, whenever the snapshot reaches b edges, enumerates all maximal cliques with more than two nodes to define hyperedges, replacing any smaller hyperedges they contain. Each hyperedge then owns one memory unit instead of each node owning one, so the model's memory scales with the number of groups rather than the number of nodes. Hypergraph convolution (Bai et al., 2021) turns these hyperedge features into node embeddings, and an injective time-decay aggregation (Equation 3) merges old hyperedge features into new larger ones, which is how several hops of neighborhood information are compressed into a single representation.","core_discovery":"HTGN's central claim is that higher-order structure can be identified cheaply from temporal locality: when edges are accumulated into short snapshots, maximal cliques within a snapshot are likely to be genuine group interactions, so they can be used as hyperedges. These hyperedges are stored in a purpose-built memory module whose features are updated by message passing and aggregated with a time-decaying injective function, and node embeddings are produced by hypergraph convolution over the hyperedges. The paper demonstrates superior Mean Reciprocal Rank on dynamic link prediction across both homogeneous and bipartite temporal graphs, and proves in Theorem 2 that HTGN distinguishes pairs of temporal computation trees that pairwise message-passing TGNNs cannot, making it strictly more expressive.","pith_inferences":["Editorial inference: The snapshot mechanism could be replaced by a sliding time window to make hyperedge construction continuous rather than batch-based, which would make HTGN applicable to irregular streams without fixed batch sizes.","Editorial inference: The memory compression idea may transfer to static graph learning, where pooling cliques into hyperedges could substitute for expensive multi-hop message passing on dense subgraphs.","Editorial inference: The ablation result that common-neighbor information still helps suggests higher-order structure and pairwise closure signals are partly complementary; a hybrid model could push accuracy further.","Editorial inference: A direct test of the expressiveness claim would be to compare HTGN with a pairwise TGNN of matched parameter count on datasets with and without known group structure, predicting the gap is large only when group structure exists."],"forward_implications":["Dynamic link prediction improves by about 5 percent MRR over the second-best method, averaged across the ten tested datasets.","GPU memory consumption drops by 30–50 percent relative to memory-based pairwise baselines like TGN and TNCN.","HTGN is strictly more expressive than pairwise message-passing temporal GNNs, meaning it can distinguish temporal computation trees those models cannot.","The same hyperedge construction approach works for both homogeneous graphs (maximal cliques) and bipartite graphs (one partition becomes hyperedges).","Memory use scales with the number of group structures rather than the number of nodes, avoiding the exponential expansion of multi-hop pairwise aggregation."],"supporting_citations":[{"why":"provides the TGN memory-module baseline and the efficiency bottleneck that HTGN addresses.","marker":"(Rossi et al., 2020)"},{"why":"supplies the hypergraph stochastic block model that the paper extends into HT-SBM for Theorem 1.","marker":"(Pister & Barthelemy, 2024)"},{"why":"frames hyperedge construction from maximal cliques and supplies Theorem 3 used in Theorem 1's proof.","marker":"(Wang & Kleinberg, 2024)"},{"why":"motivates the temporal locality of group interactions (simplicial closure).","marker":"(Benson et al., 2018)"},{"why":"defines temporal computation trees and the expressiveness notion used in Theorem 2.","marker":"(Souza et al., 2022)"},{"why":"provides the hypergraph convolution used to compute node embeddings from hyperedge features.","marker":"(Bai et al., 2021)"},{"why":"provides datasets, negative sampling, and the MRR evaluation protocol.","marker":"(Poursafaei et al., 2022)"},{"why":"supplies the Temporal Graph Benchmark datasets (tgbl-*).","marker":"(Huang et al., 2023)"},{"why":"supplies the PAC-Bayes lemma (Lemma 17) invoked in the proof of Theorem 1.","marker":"(Germain et al., 2015)"}],"fun_headline_variants":["Hyperedge-aware temporal GNN beats pairwise models, cuts memory 50%","Clique-based hyperedges improve temporal link prediction","Provably expressive temporal GNN uses hyperedges, saves memory","Group structures as hyperedges boost dynamic link prediction","Temporal GNN with hyperedges: better and 50% lighter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that maximal cliques found in short snapshots of the most recent b edges correspond to genuine group interactions, so that hyperedges built from them carry signal rather than noise for link prediction.","fun_headline_variants_meta":{"raw":{"variants":["Hyperedge-aware temporal GNN beats pairwise models, cuts memory 50%","Clique-based hyperedges improve temporal link prediction","Provably expressive temporal GNN uses hyperedges, saves memory","Group structures as hyperedges boost dynamic link prediction","Temporal GNN with hyperedges: better and 50% lighter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1369,"prompt_tokens":850,"completion_tokens":519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":434}},"tokens_in":466,"tokens_out":519,"duration_ms":5510,"temperature":1.0,"reasoning_tokens":434,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:12:10.797832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train HTGN and a strong pairwise baseline such as TGN or TNCN on a temporal network generated with no group interactions, using equal-sized embeddings; if HTGN still matches or beats the baseline, its edge is not coming from higher-order structure. Alternatively, run HTGN with snapshot sizes large enough that Theorem 1 predicts degraded hyperedge accuracy; if link prediction MRR does not drop, the hyperedge construction step is not driving the result.","supporting_citations":[],"review_version":1}