{"id":"eaf2af64-8bf4-47e6-9fe5-74119fedb513","arxiv_id":"2507.13825","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EAGLE predicts temporal links with top-k recent neighbors plus top-k shared temporal PageRank influencers, matching or beating transformer T-GNNs while running far faster.","lead":"This paper introduces EAGLE, a lightweight model for predicting future links in evolving networks. EAGLE combines a node's most recent interactions with influence scores from temporal personalized PageRank, reporting accuracy on par with or above heavier transformer models at a fraction of the runtime.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"EAGLE's headline effectiveness rests on the structure-aware score (Eq. 10), a static-PPR-inspired heuristic that is not validated on temporal graphs; if shared top-k_s T-PPR nodes do not signal future links, the hybrid collapses to the much weaker time-only module.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the structure-aware module's scoring rule, Eq. (10), is a heuristic supported only by static-graph evidence. This is indeed the most critical point because the paper's own ablations show that the time-aware module alone is far less effective on Wikipedia (69.18 vs 87.02 AP) and Reddit (31.96 vs 78.09 AP). Thus the headline claim of consistent superior performance depends on the structural term being a valid temporal signal. The static PPR experiment (Fig. 2) and the GCN-mean theorem (Thm 3.1) do not transfer to temporal graphs without additional argument: T-PPR scores evolve, top-k_s sets may be unstable, and the dot-product over shared hubs may simply encode degree effects. I also note a secondary issue: the AP/MRR values in Table 3 are inconsistent with a per-query single-positive ranking protocol (where AP would equal MRR), suggesting an ambiguous or nonstandard metric computation; this should be clarified for reproducibility, but it is not the primary concern because it would affect baselines and EAGLE similarly. The proposed test—comparing temporal T-PPR against static PPR and random top-k_s nodes, plus measuring top-k_s stability—would directly settle whether the structural mechanism is load-bearing. Given that the reader already conditioned acceptance on validating this mechanism, I agree with the CONDITIONAL verdict and see no reason to change it.","tokens_in":21157,"tokens_out":9742,"duration_ms":110795,"concrete_test":"On Wikipedia and Reddit, rank test edges by Eq. (10) using (i) temporal T-PPR (Eq. 9), (ii) static PPR on the final graph, and (iii) randomly selected top-k_s nodes. If (i) is not significantly better than (ii) and (iii) in AP/MRR, the temporal structural signal is not load-bearing. Also measure Jaccard overlap of top-k_s sets before/after inserting 1% new edges; if overlap < 0.5, the incremental T-PPR update and the structural signal are unstable.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The structure-aware module is the main source of EAGLE-Hybrid's gains: on Wikipedia, EAGLE-Time achieves AP 69.18 while EAGLE-Hybrid achieves 87.02; on Reddit, 31.96 vs 78.09. Yet Eq. (10) — the sum of products of T-PPR scores over common top-k_s nodes — is a heuristic whose only support is a static-graph experiment (Fig. 2) and a static GCN theorem (Thm 3.1). Neither establishes that, in a temporal graph, two nodes sharing high-T-PPR nodes are more likely to interact, nor that the time-decayed PPR in Eq. (9) is necessary. The score is unnormalized and can be inflated by both nodes pointing to global hubs; the top-k_s truncation may discard the informative part of the PPR distribution. Because the hybrid score (Eq. 11) adds this unbounded term to a bounded sigmoid output with only a scalar λ, the structural term can dominate regardless of the adaptive recency weights, making the model's performance depend critically on an unvalidated structural signal. If Eq. (10) is not predictive on temporal graphs, EAGLE reduces to the time-aware module and falls below DyGFormer on Wikipedia and Reddit, contradicting the abstract's 'consistently superior performance' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EAGLE, a lightweight model for temporal link prediction. It consists of a time-aware module that averages node/edge features over each node's k_r most recent neighbors, and a structure-aware module that computes a score from the sum of products of T-PPR scores over common top-k_s influential nodes. The two scores are combined in Eq. (11) with a validation-tuned scalar lambda. Experiments on seven temporal graph benchmarks compare EAGLE against JODIE, TGAT, TGN, GraphMixer, Zebra, and DyGFormer, reporting AP/MRR/HR@10 and claiming more than a 50x speedup over transformer-based T-GNNs. The paper also includes an appendix on node classification.","tokens_in":21504,"tokens_out":6150,"duration_ms":73579,"significance":"If the structural heuristic in Eq. (10) is valid on temporal graphs, the paper provides a meaningful simplifying result: a two-module model with no multi-hop message passing and no memory banks can match or outperform complex T-GNNs on standard benchmarks. The paper ships an artifact, uses widely adopted benchmarks, and includes ablations and parameter sensitivity studies. However, the central structural score is not validated on temporal graphs, and the headline efficiency claim is not documented across all seven datasets. These gaps currently prevent the significance from being fully established.","major_comments":[{"comment":"The structure-aware score is the main engine of the reported gains on Wikipedia and Reddit: EAGLE-Time achieves AP 69.18 and 31.96 while EAGLE-Hybrid achieves 87.02 and 78.09. Yet Eq. (10) is supported only by a static-graph PPR experiment (Fig. 2) and a static GCN theorem (Thm. 3.1). Neither support establishes that, on a temporal graph, two nodes sharing top-k_s T-PPR nodes are more likely to interact, nor that the product form is appropriate. Please provide a temporal validation of Eq. (10): for example, compare EAGLE-Struc against static-PPR and common-neighbor baselines on the temporal benchmarks, and report the predictive power of the structural term alone with negative controls. Without this, the 'consistently superior' claim rests on an unvalidated heuristic.","section":"§3.2.2, Eq. (10)"},{"comment":"The hybrid score adds an unbounded structural term to a bounded sigmoid time score, and the only adaptation is a validation-tuned scalar lambda multiplied by exp(-tbar_v)+exp(-tbar_u). This mechanism cannot downweight the structural term when it is uninformative, and the structural term itself has no recency adaptivity. Please report the selected lambda values per dataset and provide ablation results with normalized structural scores (e.g., rank-based or min-max scaling) to support the claim that the model 'balances' the two signals.","section":"§3.2.3, Eq. (11)"},{"comment":"The abstract's claim of 'more than 50x speedup over effective transformer-based T-GNNs' is not substantiated. Table 4 reports efficiency for only three of the seven datasets (Contacts, Wikipedia, WikiTalk), and DyGFormer either OOMs or times out on two of them; the only completed comparison is Wikipedia, where the speedup is about 91x. Please report efficiency for all datasets on which baselines complete, and state the exact comparison underlying the 50x claim. In addition, Algorithm 1 delegates the T-PPR update to the black-box procedure T-PPR_UPDATING from [25]; because the stated O(n k_s log k_s) update and O(k_s |V|) memory are imported from that reference, the paper should specify or implement the update, or explicitly bound the claims by the cited results.","section":"§4.2.2, Table 4"},{"comment":"Several 'superior' differences are within one standard deviation of the best baseline: for example, Wikipedia AP is 87.02±0.29 for EAGLE-Hybrid versus 86.96±0.22 for DyGFormer, and Reddit AP is 78.09±0.66 versus 77.92±0.33. To support 'consistently superior performance', please report the number of seeds and run statistical significance tests or provide a clear statement of which differences are significant.","section":"§4.2.1, Table 3"}],"minor_comments":[{"comment":"The case where the intersection of the top-k_s T-PPR sets is empty is not discussed; if the structural score is zero for many test pairs, ranking behavior should be analyzed. Please report the fraction of test pairs with a non-empty intersection.","section":"§3.2.2"},{"comment":"The concatenation [h^r_v(t), h^r_v(t)] appears to be a typo for [h^r_v(t), h^s_v(t)], since the structure-aware representation is otherwise unused in the classification head.","section":"Appendix A.1, Eq. (14)"},{"comment":"Efficiency is omitted for AskUbuntu, SuperUser, LastFM, and Reddit; please add these results or explain why they are not reported.","section":"§4.2.2, Table 4"},{"comment":"The statement that EAGLE-Struc is 'training-free' is misleading because alpha, beta, and k_s are tuned on validation data; clarify that 'training-free' means no gradient-based optimization.","section":"§3.2.4"},{"comment":"The motivational experiment does not state which base model is used; please specify the predictor and the number of repeats so the observation is reproducible.","section":"§3.1.1, Figure 1"},{"comment":"The phrase 'adaptive weighting mechanism' is stronger than what is implemented: lambda is a single scalar tuned on the validation set, so the mechanism is static after validation. Consider describing it as a validation-tuned trade-off parameter.","section":"§3.2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the VLDB audience insofar as it emphasizes efficiency, but the efficiency evaluation is incomplete and the T-PPR update is reused from the authors' prior Zebra work. The novelty is the hybrid scoring scheme, so the manuscript should more clearly separate its contributions from Zebra and provide the missing temporal validation of Eq. (10). If the authors can supply that validation and complete efficiency measurements, the result would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"EAGLE is a refreshingly simple temporal link predictor that, if the numbers hold up, should become a baseline people actually run. The combination of top-k recent neighbors and a T-PPR-derived shared-influence score is new, and the efficiency numbers are real: a two-layer MLP over k_r + k_s neighbors is indeed cheaper than transformer T-GNNs.\n\nThe paper does the right empirical work: seven datasets, ablations, sensitivity, node classification, and an artifact URL. The structural module is not merely decorative. EAGLE-Struc, which is exactly Eq (10) with no training, reaches 86.08 AP on Wikipedia and 74.85 on Reddit, close to DyGFormer. That is direct temporal-graph evidence for the structural heuristic, not just the static-graph motivation. The stress-test concern about Eq (10) being unsupported on temporal graphs overlooks that this variant is itself a temporal-graph experiment.\n\nThat said, there are real soft spots. The hybrid score in Eq (11) is a hand-designed, unnormalized sum of a bounded sigmoid and an unbounded structural term; the 'adaptive' weighting is a per-dataset scalar lambda plus a recency decay, not a learned mechanism. The 50x speedup claim in the abstract is supported in the main tables only on Wikipedia among transformer baselines that actually finish—elsewhere DyGFormer is TLE or OOM, so the comparison is narrower than the wording suggests. The artifact is a bare URL with no commit hash, and the metric protocol, particularly how negative candidates enter the AP computation, should be stated more carefully.\n\nNone of this undermines the central result: a simple model that is fast and competitive on standard benchmarks is useful. The structural heuristic is empirically validated by the Struc variant itself, so the 'collapse to time-only' worry is overstated. The paper deserves a serious referee; I would lean conditional acceptance, with requests for precise speedup claims, a runnable artifact, and a small analysis of when Eq (10) helps versus hurts (AskUbuntu is an obvious case).","headline":"EAGLE is a genuinely simple and fast temporal link predictor whose structural module is more empirically grounded than the stress-test suggests; the main soft spots are overbroad speedup claims and a hand-tuned hybrid with a fixed lambda.","tokens_in":22027,"tokens_out":3310,"would_cite":true,"duration_ms":43623,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that temporal link prediction needs no complex T-GNN: a hybrid of a recent-neighbor score and a shared-influence PageRank score matches or beats state-of-the-art models while running over 50x faster.","keywords":["temporal link prediction","temporal graph neural networks","personalized PageRank","dynamic graphs","recent neighbor aggregation","efficient graph learning","link prediction","graph representation learning"],"falsifier":"Take any benchmark graph and keep the set of top-$k_s$ nodes fixed but replace their T-PPR scores in Equation (10) with random values drawn from the same range. If EAGLE-Hybrid's AP/MRR drops only a little, the structural module's contribution is not the T-PPR ranking itself; if it collapses, the specific T-PPR values are doing the work. A complementary test: on a graph with strong recency-driven links and no recurring hubs, EAGLE-Hybrid should reduce to EAGLE-Time via the adaptive weight; if the structural term still hurts, the weight mechanism is over-trusting structure.","tokens_in":1793,"feed_emoji":"⚡","tokens_out":1935,"duration_ms":89485,"temperature":0.7,"pith_summary":"The paper tries to establish that the two attributes that actually matter for temporal link prediction are a node's most recent neighbors and the globally influential nodes it shares with a candidate partner, and that a model built from just these two signals can beat complicated temporal graph neural networks. EAGLE combines a time-aware module, which averages features of a node's $k_r$ most recent neighbors, with a structure-aware module that sums products of temporal personalized PageRank scores over shared top-$k_s$ influential nodes. An adaptive weighting term shifts between the two depending on how recently each node has interacted, so sparse or stale graphs fall back on structural evidence. On seven real-world temporal graphs the full model consistently matches or outperforms state-of-the-art T-GNNs, and the paper reports a more than 50x speedup over transformer-based T-GNNs. If true, this means the expensive machinery of multi-hop message passing, memory banks, and transformer encoders is not necessary for strong temporal link prediction.","feed_headline":"50x faster temporal link prediction with two simple scores","feed_subtitle":"EAGLE matches or beats transformer T-GNNs on seven benchmarks using recent neighbors and shared PageRank hubs.","key_machinery":"The load-bearing object is the structure-aware score in Equation (10): for a candidate pair $(v,u)$, sum $\\pi_v(t)[v_i] \\cdot \\pi_u(t)[v_i]$ over the shared nodes in the top-$k_s$ entries of both endpoints' temporal personalized PageRank vectors. It is the paper's proxy for long-term global structural dependence, and it carries the whole 'global patterns' half of the argument. This score is training-free and deterministic, can be updated incrementally as new events arrive, and is motivated by a cited theorem that multi-hop GCN propagation concentrates on paths that PPR ranks. The time-aware module, by contrast, is the learned half: a two-layer MLP over the averaged features of each node's $k_r$ most recent neighbors. The adaptive weight in Equation (11) is what decides, per node pair, how much of the final score comes from each side.","core_discovery":"EAGLE's central discovery is that a hybrid of two simple scoring functions is sufficient for state-of-the-art temporal link prediction. The time-aware score represents each endpoint by averaging the concatenated node and edge features of its $k_r$ most recent neighbors, then passes the concatenation of the two endpoint representations through a two-layer MLP. The structure-aware score is computed between endpoints $v$ and $u$ as the sum, over nodes that appear in both of their top-$k_s$ temporal personalized PageRank vectors, of $\\pi_v(t)[v_i]\\cdot \\pi_u(t)[v_i]$, with no learned parameters. The final hybrid score weights the time-aware score by $\\exp(-\\bar{t}_v)+\\exp(-\\bar{t}_u)$ and adds the structural score. Across seven datasets, the hybrid beats each module alone, and the paper reports effectiveness at least matching existing T-GNNs while training and inference become dramatically faster and lighter.","pith_inferences":["Beyond the paper: on graphs with stable communities and long timespans, I expect the structure-aware score from Equation (10) to contribute most, while on bursty recency-driven graphs the adaptive weight should down-weight it; the seven benchmarks do not isolate this axis explicitly.","Beyond the paper: the same T-PPR shared-influence score could serve as a cheap candidate generator or pre-filter ahead of a heavier reranker, since it is training-free and deterministic.","Beyond the paper: an ablation that replaces the T-PPR values in Equation (10) with binary indicators of shared top-$k_s$ nodes would test whether it is the ranking or the magnitudes that matter; the paper does not run this variant."],"forward_implications":["Inference for a node pair touches only $k_r$ most recent neighbors and $k_s$ top T-PPR nodes, so runtime grows with $k_r + k_s$ rather than with multi-hop neighborhoods or a full transformer context.","The structure-aware score is training-free, so the hybrid can be deployed with almost no learned parameters beyond a two-layer MLP for the time-aware module.","On datasets where recent interactions dominate, the adaptive weight automatically leans on the time-aware score; where interactions are sparse or old, it leans on the structural score.","Because no memory banks or multi-hop message passing are needed, training time and peak GPU memory are dramatically lower, which the paper reports as over 50x speedup versus transformer-based T-GNNs.","The same two representations extend to node classification, where the paper reports large gains and speedups over the same baselines in the appendix."],"supporting_citations":[{"why":"Supplies the temporal personalized PageRank (T-PPR) algorithm and the incremental update procedure that EAGLE's structure-aware module relies on.","marker":"[25]"},{"why":"Provides the theorem linking multi-hop GCN propagation to path probabilities, which motivates using PPR-ranked influential nodes as the structural signal.","marker":"[56]"},{"why":"Used in the paper's static-graph PPR experiment, where APPNP shows that retaining only the top-$k_s$ PPR neighbors is sufficient for link prediction accuracy.","marker":"[12]"},{"why":"Provides the Contacts and LastFM temporal graph datasets used in the main evaluation.","marker":"[40]"},{"why":"Provides the Wikipedia and Reddit temporal interaction datasets used for training and evaluation.","marker":"[20]"},{"why":"Provides the AskUbuntu, SuperUser, and Wiki-Talk temporal graph datasets used in the main evaluation.","marker":"[37]"},{"why":"An efficiency-focused temporal GNN baseline whose claim that complicated architectures may be unnecessary EAGLE extends and compares against.","marker":"[6]"},{"why":"The transformer-based temporal GNN baseline over which EAGLE reports the more than 50x speedup and effectiveness comparisons.","marker":"[59]"},{"why":"Supplies the evaluation protocol and the node-classification benchmark datasets used in the appendix.","marker":"[19]"}],"fun_headline_variants":["Two simple scores match T-GNNs at 50x speed","EAGLE: fast temporal link prediction with a neat hybrid","50x faster, same accuracy: a two-score recipe for dynamic graphs","Drop message passing: hybrid recency + PageRank wins"],"cache_read_input_tokens":24064,"weakest_assumption_plain":"The hybrid's effectiveness rests on the premise that two nodes are more likely to link when they share many of the same globally influential neighbors, with influence measured by temporal personalized PageRank; if that shared-influence score does not track future links in evolving graphs, the whole structural half of the model fails.","fun_headline_variants_meta":{"raw":{"variants":["Two simple scores match T-GNNs at 50x speed","EAGLE: fast temporal link prediction with a neat hybrid","50x faster, same accuracy: a two-score recipe for dynamic graphs","Drop message passing: hybrid recency + PageRank wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2445,"prompt_tokens":948,"completion_tokens":1497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1424}},"tokens_in":564,"tokens_out":1497,"duration_ms":12179,"temperature":1.0,"reasoning_tokens":1424,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:16:03.657359+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any benchmark graph and keep the set of top-$k_s$ nodes fixed but replace their T-PPR scores in Equation (10) with random values drawn from the same range. If EAGLE-Hybrid's AP/MRR drops only a little, the structural module's contribution is not the T-PPR ranking itself; if it collapses, the specific T-PPR values are doing the work. A complementary test: on a graph with strong recency-driven links and no recurring hubs, EAGLE-Hybrid should reduce to EAGLE-Time via the adaptive weight; if the structural term still hurts, the weight mechanism is over-trusting structure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the temporal personalized PageRank (T-PPR) algorithm and the incremental update procedure that EAGLE's structure-aware module relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the theorem linking multi-hop GCN propagation to path probabilities, which motivates using PPR-ranked influential nodes as the structural signal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Wikipedia and Reddit temporal interaction datasets used for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An efficiency-focused temporal GNN baseline whose claim that complicated architectures may be unnecessary EAGLE extends and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The transformer-based temporal GNN baseline over which EAGLE reports the more than 50x speedup and effectiveness comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the evaluation protocol and the node-classification benchmark datasets used in the appendix."}],"review_version":1}