{"id":"ff3b115a-c166-4fc0-92fd-47a020d1afd4","arxiv_id":"2411.10957","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"IMPaCT reweights graph message passing to align first and second moments across chronological train/test splits, reporting accuracy gains on ogbn-mag and a theoretical W1 bound.","lead":"This paper proposes IMPaCT, a set of message passing corrections that align the distribution of node representations between old and new time periods in graphs split chronologically. The method reports accuracy gains on citation graphs and comes with a theoretical generalization bound, though the proof has gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5's contraction is unsupported: Lemma 4 only proves a strict < S D bound, not the uniform (S−δ)D gap needed for G^(k)>1, so the generalization bound does not follow.","rationale":"The reader's stated weakest assumption is Assumption 3, but the reader's rationale also notes that the proof of Lemma 4 does not establish the contraction needed for Theorem 4.5. My analysis confirms that this is the most load-bearing flaw: it is an internal mathematical gap in the paper's central theoretical claim, independent of empirical questions about Assumption 3. Even if Assumption 3 held perfectly, Theorem 4.5 would still be unsupported because Lemma 4's proof produces only '< S D', not the uniform '(S−δ)D' that is needed to make G^(k)>1. The concrete counterexample with two point masses shows why the gap cannot be inferred from the given inequalities. This concern does not depend on whether real-world temporal graphs satisfy the modeling assumptions, so it is more decisive than the Assumption 3 verification issue. The empirical comparisons also merit criticism, but the theoretical result is the paper's headline contribution and it fails at a specific derivational step. I therefore agree with the reader's REJECT verdict, though through a different primary route; no verdict adjustment is needed.","tokens_in":33848,"tokens_out":3976,"duration_ms":44756,"concrete_test":"Independently re-derive Lemma 4 with n=2, S=1, μ_1=δ_0, μ_2=δ_D, and weights η=(1−ε,ε), ν=(ε,1−ε). The W1 distance between the two mixtures is (1−2ε)D; for any positive δ, choosing ε<δ/2 yields a distance larger than (1−δ)D, directly contradicting the uniform (S−δ)D claim of Lemma 4. If the authors cannot replace Lemma 4 with a version whose gap is bounded below by a positive constant derived from the actual PMP weights, Theorem 4.5 and the generalization-error bound should be withdrawn or substantially revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim rests on Lemma 4 (Appendix A.8.1), which is then used to prove Theorem 4.5 and Corollary 4.5.1. Lemma 4 asserts that for distributions μ_i with pairwise W1 ≤ D and weights η_i, ν_i ∈ (0, S) with equal total S, one has W1(Ση_i μ_i, Σν_i μ_i) < (S−δ)D for some positive δ. The proof, however, establishes only the chain ≤ Σ_{i:δ_i≥0} δ_i D < Σ_{i:δ_i≥0} η_i D < S D. A strict upper bound of S D does not imply a uniform positive gap δ; the gap can vanish as weights approach 0 or S. Concretely, take S=1, μ_1=δ_0, μ_2=δ_D, η=(1−ε,ε), and ν=(ε,1−ε). Then W1(μ_1,μ_2)=D and W1(Ση_i μ_i, Σν_i μ_i)=(1−2ε)D. For any proposed δ>0, choosing ε<δ/2 makes the distance exceed (1−δ)D, so no fixed δ works. Theorem 4.5 needs G^(k)=1/(1−ε)>1 with ε = min_{y,t,t'} ε_{y,t,t'} > 0, but no lower bound on ε is established. Thus even accepting Assumption 3, the advertised layer-by-layer contraction and the resulting upper bound on generalization error are not derived from the presented proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses chronological-split node classification by proposing IMPaCT, a set of message passing modifications (MMP, PMP, GenPMP, PNY, JJNORM) that aim to make the first and second moments of aggregated node representations invariant to node time. The authors introduce Assumptions 1-4 on label/feature stationarity and the factorization of neighbor-label-time connectivity, prove moment-invariance theorems and a claimed Wasserstein-1 contraction bound leading to a generalization-error upper bound, propose a synthetic Temporal Stochastic Block Model (TSBM), and report experiments on ogbn-mag, ogbn-arxiv, ogbn-papers100m, and synthetic graphs.","tokens_in":34177,"tokens_out":4966,"duration_ms":58194,"significance":"If the theoretical claims were correct, the paper would contribute a scalable preprocessing approach to a practically important but underexplored problem, with a relatively rare theoretical analysis connecting chronological splits to Wasserstein-1 contraction in GNNs. The empirical improvements on ogbn-mag (0.9178 test accuracy versus 0.8789 for the LDHGNN baseline) and the TSBM testbed are useful, and the complexity analysis in Table 4 is a strength. However, the central contraction theorem is not proven, so the advertised upper bound on generalization error does not follow from the presented arguments.","major_comments":[{"comment":"The proof of Lemma 4 establishes only W1(Ση_i μ_i, Σν_i μ_i) ≤ Σ_{i:η_i≥ν_i}(η_i−ν_i)D < S D. A strict inequality < S D does not yield the stated existence of a positive δ with W1 ≤ (S−δ)D; the gap may vanish. For example, with S=1, μ_1=δ_0, μ_2=δ_D, η=(1−ε,ε), and ν=(ε,1−ε), the W1 distance is (1−2ε)D, so for any proposed δ>0 one can choose ε<δ/2 to violate the conclusion. Thus the lemma as stated is false.","section":"Appendix A.8.1, Lemma 4"},{"comment":"The proof of Theorem 4.5 uses Eqs. (90)-(93) to define G^(k)=1/(1−ϵ)>1, but ϵ is defined as min_{y,t,t'} ϵ_{ytt'}, and no lower bound on the individual ϵ_{y˜y t t'} is proven. Since Lemma 4 does not provide a uniform positive gap, the claimed contraction W1(m(k+1)_yt, m(k+1)_yt') ≤ (G/G^(k))W with G^(k)>1 is not established. Corollary 4.5.1 and the associated generalization-error claim rest directly on this unsupported step.","section":"Section 4.3, Theorem 4.5 / Appendix A.8.4"},{"comment":"All moment-invariance theorems (4.1, 4.2, 5.1, 5.2) and the W1 analysis depend on Assumption 3, Eq. (6). The paper supports this assumption only by plotting unscaled relative connectivity and stating that the shapes look similar, and it explicitly acknowledges that the analysis is not a formal proof. A quantitative goodness-of-fit test or a robustness analysis is needed before the theory can be considered validated for real graphs.","section":"Appendix A.1.2, Assumption 3"},{"comment":"The claimed upper bound of the generalization error is introduced by stating Eq. (3) without the assumptions required by the cited domain adaptation results (e.g., a symmetric hypothesis class and a specified loss function). The paper does not instantiate these conditions for the actual multi-class node classification setting, so even if the W1 contraction were established, the link to risk would still not be a rigorous derivation.","section":"Section 2.2, Eq. (3)"}],"minor_comments":[{"comment":"The column header 'Baseline MMP' is ambiguous; it obscures that MMP underperforms the baseline in the reported synthetic experiments. Please reformat the table and discuss this result explicitly.","section":"Table 3"},{"comment":"The citation for LDHGNN is given as (Wang, 2024) in the text, while the reference list entry appears as Wang, Yili; please make the citation style consistent.","section":"Section 6.2"},{"comment":"The dataset name is misspelled as 'obgn-mag'; it should be 'ogbn-mag'.","section":"Table 6"},{"comment":"The estimator ^α_t is written with a numerator term that is not obviously nonnegative; please clarify the derivation and state whether any clipping or lower-bounding is applied in the implementation.","section":"Eq. (20)"},{"comment":"The notation µ^JJ_M(y,t) and µ^JJ_M(·,t) is used before it is clearly introduced; please standardize the definitions and their reference in the surrounding text.","section":"Definition 5.3"}],"recommendation":"reject","confidential_remarks":"The empirical results and the TSBM testbed may be of interest, but the advertised theoretical contribution is unsupported by the current proof of Lemma 4 and Theorem 4.5. Because the contraction and generalization bounds are central to the paper's stated contributions, acceptance does not seem possible without introducing and proving a genuinely different contraction result, which is beyond a routine revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely useful: reweight message-passing contributions to equalize the temporal-neighbor distributions under the chronological split. The PMP/MMP/GenPMP family is well explained, and the first-moment invariance theorems (4.1, 4.2) do hold under Assumption 3. The TSBM synthetic generator is a nice, reusable controlled testbed, and the synthetic experiments are thorough and show consistent gains across SGC and GCN. Credit is due for those pieces.\n\nThe soft spots are real and one is load-bearing. The stress-test note is correct: Lemma 4's proof establishes W1 < S D, not a uniform (S – delta)D gap. The gap can vanish as the weights approach 0 or S, so the layer-wise contraction in Theorem 4.5 is unsupported. The generalization bound therefore does not follow from the presented math. This is not a minor gap; it is the advertised theoretical contribution.\n\nTwo empirical issues also need flagging. The abstract's \"3.8% over SOTA\" is actually against the LDHGNN baseline in Table 2, not a real SOTA comparison, and no prior invariant GNN methods (EERM, SR-GNN) are included as baselines. Also, JJNORM fits normalization constants to test-node statistics. That is acceptable in a transductive setting but weakens the \"generalization\" framing.\n\nTo be fair, the paper is honest that Assumption 3's support is visual, not a formal test, and the limitations section is measured. But the main theorem is broken, and the empirical headline is overstated. The idea and TSBM are salvageable; the theory needs either a real fix or explicit removal from the claims.\n\nI would send this to peer review because the method and testbed deserve referee time, but with the expectation of major revision: correct or delete the contraction theorem, rerun against actual SOTA and prior invariant methods, and clarify what the empirical gains are relative to.","headline":"A clearly motivated temporal-invariance trick worth knowing about, but the advertised generalization bound is not proven and the SOTA comparison is over the authors' own baseline.","tokens_in":34761,"tokens_out":2017,"would_cite":false,"duration_ms":25330,"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":"By reweighting message passing according to timestamp symmetry, IMPaCT makes node representations time-invariant and bounds the Wasserstein distance between train and test distributions, achieving a 3.8-point gain on ogbn-mag.","keywords":["temporal graphs","chronological split","domain adaptation","invariant message passing","moment alignment","Wasserstein-1 distance","semi-supervised node classification","graph neural networks"],"falsifier":"On a real chronological graph with observed labels, estimate the relative connectivity $P_{y,t}(\\tilde{y},\\tilde{t})$ for a fixed label pair and lag $\\Delta$, and test whether the ratio $P_{y,t}(\\tilde{y},t+\\Delta)/P_{y,t'}(\\tilde{y},t'+\\Delta)$ is constant in $t,t'$ after scaling. If a quantitative goodness-of-fit test of the product form $f(y,t)\\,g(y,\\tilde{y},|\\tilde{t}-t|)$ shows systematic residuals correlated with target time, Assumption 3 is false and the moment-invariance theorems no longer apply.","tokens_in":33574,"feed_emoji":"📈","tokens_out":10589,"duration_ms":99770,"temperature":0.7,"pith_summary":"In a graph whose train and test nodes are separated by time, the paper claims that distribution shift can be removed by correcting the way each node averages its neighbors' messages. It introduces IMPaCT, a family of message-passing rewrites: PMP and MMP adjust first-moment (mean) invariance by duplicating or dropping edges based on timestamp symmetry, while PNY and JJNORM align second-moment (variance) statistics. Under a factorized model of neighbor connectivity, PMP makes the Wasserstein distance between train and test message distributions contract layer by layer, bounding the generalization error. On the ogbn-mag citation graph, PMP plus JJNORM raises test accuracy by 3.8 percentage points over the LDHGNN baseline. If correct, this gives a preprocessing recipe for chronological-split GNNs that works without knowing test labels.","feed_headline":"Edge reweighting wins 3.8 points on a temporal citation graph","feed_subtitle":"Reweighting neighbor messages by time symmetry prevents distribution shift; test accuracy rises 3.8 points.","key_machinery":"The load-bearing object is Assumption 3: $P_{y,t}(\\tilde{y},\\tilde{t}) = f(y,t)\\,g(y,\\tilde{y},|\\tilde{t}-t|)$, stating that neighbor connectivity is a product of a target-time scale and a time-lag decay. PMP exploits the identity that after doubling one-sided neighbor classes, the sum over neighbor times becomes $2f(y,t)\\sum_{\\tau\\ge0} g(y,\\tilde{y},\\tau)$, so the target-time dependence cancels in the weighted average; this cancellation is what preserves first-moment invariance layer by layer. JJNORM uses a corollary of the same identity: under an additional label-independent decay assumption (Assumption 4), the covariance of aggregated messages at time $t$ differs from that at $t_{\\max}$ only by a scalar $\\alpha_t^2$, allowing a cheap unbiased normalization. The Wasserstein contraction theorem (Theorem 4.5) feeds this invariance into the domain-adaptation bound $R_{te} \\le R_{tr} + W_1$.","core_discovery":"The central discovery is that the distribution shift in chronological splits comes from an asymmetry in which past and future neighbors a node can see: a node near the end of the timeline receives neighbors from both sides while boundary nodes cannot, so ordinary averaging message passing gives different effective weight to the same time lag for different target times. IMPaCT's PMP fixes this by doubling the weight of neighbor timesteps that occur on only one side, making the effective aggregation weights depend only on label and absolute time difference. The paper proves the first moment of the aggregated message then stays invariant from layer to layer, and derives Wasserstein-1 bounds: $O(C^{1/3}V^{1/3})$ generally and $O(\\tau\\sqrt{\\log C})$ under sub-Gaussian representations, with a layer-wise contraction factor $G^{(k)} > 1$. For the second moment, JJNORM scales each node's final message by a per-time constant $\\alpha_t$, estimated without test labels, so that covariance matrices align to the test time. Empirically, on ogbn-mag, LDHGNN+PMP+JJNORM reaches 0.9178 test accuracy versus 0.8789 for the LDHGNN baseline, and on ogbn-arxiv GenPMP adds 4.0 points to a linearized baseline.","pith_inferences":["Beyond the paper, the same one-sided/two-sided neighbor asymmetry should appear in any temporally ordered graph, so PMP-style edge reweighting is a candidate preprocessing step for future-dated link prediction and temporal recommendation.","The paper's visual support for Assumption 3 could be made quantitative; a reader could fit $f$ and $g$ to measured connectivity and check whether residuals correlate with target time, a test not reported here.","Beyond the paper, the layer-wise contraction result suggests a design principle: any aggregation whose weights depend only on label and absolute time lag will inherit the bound, so other reweighting schemes such as attention over time lags should be explored.","JJNORM's per-time scalar $\\alpha_t$ could be applied as a test-time feature standardization even when the base GNN is not decoupled, though the paper only recommends it for decoupled GNNs at scale."],"forward_implications":["A user of any averaging-based spatial GNN can apply PMP to a chronological split without changing the model, by duplicating edges whose neighbor timestamps are one-sided relative to the target node.","When the assumptions hold, the train/test gap in final representations is provably bounded and shrinks with network depth rather than growing.","On ogbn-mag, PMP+JJNORM beats the reported baseline by 3.8 percentage points, and PMP alone still improves it, so the correction works without test labels.","GenPMP extends the correction to datasets where node counts per timestamp are highly uneven, where plain PMP's approximation degrades.","For decoupled GNNs, all IMPaCT operations occur during preprocessing, adding no training-time cost beyond the original message-passing procedure."],"supporting_citations":[{"why":"Supplies the classic domain-adaptation bound that motivates matching train and test distributions.","marker":"Ben-David et al., 2006; 2010"},{"why":"Establishes the Wasserstein-1 domain-adaptation bound that the paper's Theorem 4.5 targets.","marker":"Redko et al., 2017"},{"why":"Provides the stochastic block model that TSBM extends to generate synthetic temporal graphs satisfying the assumptions.","marker":"Holland et al., 1983"},{"why":"Defines averaging message passing and serves as the general spatial GNN baseline for synthetic experiments.","marker":"Kipf & Welling, 2017"},{"why":"SGC is the decoupled baseline on synthetic graphs where IMPaCT methods are first tested.","marker":"Wu et al., 2019"},{"why":"LDHGNN is the state-of-the-art baseline on ogbn-mag against which PMP+JJNORM is compared.","marker":"Wang, 2024"},{"why":"Open Graph Benchmark supplies ogbn-mag, ogbn-arxiv, and ogbn-papers100m with chronological splits used in the experiments.","marker":"Hu et al., 2020"}],"fun_headline_variants":["Time-symmetric message passing beats temporal graph shift","Chronological split GNN: 3.8-point gain with invariance","Provable invariant GNN for chronological splits is SOTA","Double one-sided neighbor weights to fix temporal shift","Temporal graph domain shift solved by invariant messages"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Assumption 3: the distribution of a node's neighbors' labels and times factors as $P_{y,t}(\\tilde{y},\\tilde{t}) = f(y,t)\\,g(y,\\tilde{y},|\\tilde{t}-t|)$; if real graphs do not factor this way, the moment-alignment theorems and the Wasserstein bound do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Time-symmetric message passing beats temporal graph shift","Chronological split GNN: 3.8-point gain with invariance","Provable invariant GNN for chronological splits is SOTA","Double one-sided neighbor weights to fix temporal shift","Temporal graph domain shift solved by invariant messages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3544,"prompt_tokens":1045,"completion_tokens":2499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2420}},"tokens_in":661,"tokens_out":2499,"duration_ms":21822,"temperature":1.0,"reasoning_tokens":2420,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:06:24.542714+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a real chronological graph with observed labels, estimate the relative connectivity $P_{y,t}(\\tilde{y},\\tilde{t})$ for a fixed label pair and lag $\\Delta$, and test whether the ratio $P_{y,t}(\\tilde{y},t+\\Delta)/P_{y,t'}(\\tilde{y},t'+\\Delta)$ is constant in $t,t'$ after scaling. If a quantitative goodness-of-fit test of the product form $f(y,t)\\,g(y,\\tilde{y},|\\tilde{t}-t|)$ shows systematic residuals correlated with target time, Assumption 3 is false and the moment-invariance theorems no longer apply.","supporting_citations":[{"cited_title":"Theoretical analysis of domain adaptation with optimal transport","cited_arxiv_id":null,"evidence_quote":"Establishes the Wasserstein-1 domain-adaptation bound that the paper's Theorem 4.5 targets."}],"review_version":1}