{"id":"5d3fad39-5340-4bbf-b963-2e0a3ea4ad19","arxiv_id":"2411.13821","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"CausalMP rewires a graph using conditional-entropy asymmetry and mutual information, then trains a GNN on the rewired graph to improve link prediction and few-shot node classification.","lead":"This paper introduces CausalMP, a graph neural network that rewrites edges based on an entropy-based dependency score and reports better link prediction and node classification on both homophilic and heterophilic graphs. A generalist reader might care because it offers a new way to separate useful from misleading connections in networks where connected nodes often belong to different classes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The δH estimator in Definition 3.4 cannot establish causal direction: it is an absolute value of a conditional-entropy difference that also responds to unequal marginal entropies, so the causal-structure claim is unsupported and the gains may come from generic rewiring.","rationale":"The paper's central claim has two parts: explicit causal structure discovery and superior link prediction. The second is reasonably supported by Tables 1–4, though code and threshold values are absent. The first is not supported. The reader flagged the same weakest assumption, and the stress test shows the flaw is concrete: absolute conditional-entropy difference is not a causal orientation rule and is sensitive to marginal entropy imbalance even under independence. A synthetic identifiability check would settle it. I do not recommend outright rejection because the augmentation may still be useful empirically; however, the causal interpretation must be either justified with explicit identifiability assumptions and validation on known SCMs, or removed/weakened, and a non-causal control (e.g., random or marginal-entropy-based rewiring) should be included. This keeps the reader's CONDITIONAL verdict, with an additional mandatory condition.","tokens_in":16194,"tokens_out":5746,"duration_ms":57670,"concrete_test":"Simulate a set of node pairs with known ground-truth data-generating processes: (a) X→Y via additive noise, (b) Y→X, and (c) X⊥Y with unequal marginal entropies. Feed the simulated feature matrices through the exact KDE and Eq. (11) pipeline of §4.1, and record the sign and magnitude of δH together with whether the pruning threshold μH+λ1σH fires. If orientation accuracy is near chance or if null independent pairs trigger pruning at a substantial rate, the estimator is not recovering causal direction or heterophily, and the central claim should be withdrawn or restricted to 'generic topology modification'.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Definition 3.4 is the load-bearing step: δH(i,j)=|H(X_j|X_i)−H(X_i|X_j)| is claimed to reveal cause-effect direction and heterophily, and the pruning rule in §4.1 reorients edges based on this score. This fails as a causal estimator for two reasons. First, the absolute value erases the sign needed to orient an edge, and conditional-entropy asymmetry is not a causal identifiability criterion: without structural assumptions (e.g., additive noise, non-Gaussianity, or invariance across environments), H(X_j|X_i)>H(X_i|X_j) does not imply X_i→X_j. The Gaussian-noise imputation in §4.1 perturbs features but is not a do-type intervention on the data-generating mechanism, so it does not supply the missing direction. Second, δH conflates dependency with marginal entropy imbalance: for independent X,Y, H(X|Y)=H(X) and H(Y|X)=H(Y), so δH=|H(X)−H(Y)|, which can be arbitrarily large. The 'insight' in Eq. (12) does not fix this; its derivation omits conditioning on N− in H(N+|x), and the subsequent text asserts rather than proves the heterophily link. The Texas case study shows only that the score is higher on heterophilic edges; it never validates direction. Thus the central causal claim is unsupported, and the Table 4 gains could be generic augmentation effects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CausalMP, a method for heterophilic graph learning that iteratively rewires the input graph based on a claimed causal structure estimated from conditional-entropy asymmetries between connected nodes. The authors formulate the problem in a Bayesian experimental-design framework, define a cause-effect significance score δH (Eq. 11), and use it to orient or prune edges and to add edges via mutual information. They report superior link prediction on nine datasets (Table 1) and consistent but modest node-classification gains when training base models on the learned structure (Table 2). The paper includes ablations of the loss coefficients, the intervention ratio, and the edge-modification components, plus a case study on Texas.","tokens_in":16559,"tokens_out":6040,"duration_ms":51592,"significance":"If the causal interpretation were justified, the paper would offer a novel bridge between causal discovery and heterophilic GNN design, with a potentially generalizable way to detect asymmetric dependencies. The empirical results are reasonably strong: CausalMP outperforms all baselines on link prediction on both homophilic and heterophilic graphs, and the node-classification gains are consistent across six base models, with ablations indicating that both the δH-based rewiring and the MI-based edge addition contribute. The complexity analysis is useful, and the case study provides some evidence that δH is higher on label-heterophilic edges. However, the central conceptual claim—that the method learns an explicit causal structure—is not substantiated, and the theoretical derivation from the Bayesian objective to the δH estimator is asserted rather than derived.","major_comments":[{"comment":"The causal-identification claim is load-bearing and unsupported. Definition 3.4 defines δH(i,j) as the absolute value |H(X_j|X_i) − H(X_i|X_j)|, yet the text immediately states that 'X_j is a cause of X_i if it is positive inside the absolute-value sign'; the absolute value erases the sign needed to orient an edge. More fundamentally, conditional-entropy asymmetry is not a causal identifiability criterion without additional structural assumptions (e.g., additive noise, non-Gaussianity, or invariance across environments). For independent variables X,Y, δH = |H(X) − H(Y)|, which can be arbitrarily large, so the score conflates dependency with marginal-entropy imbalance rather than isolating cause-effect. Because the pruning and edge-orientation rules in §4.1 rely entirely on this score, the claimed causal structure A_c is not established. I recommend a concrete test on synthetic data with known ground-truth causal directions (e.g., additive-noise models) and on independent variables with unequal entropies, to show whether δH recovers orientation or merely asymmetry.","section":"§3.2, Eq. (11), Definition 3.4; §4.1"},{"comment":"The analytical justification for linking δH to heterophily is not valid as written. Eq. (12) states H(N_i^-|x_i) + H(x_i|N_i^-) = H(x_i,N_i^+,N_i^-) − H(N_i^+|x_i) − I(x_i,N_i^-). This identity does not follow from standard entropy relations, and H(N_i^+|x_i) is not conditioned on N_i^- as it should be if the joint entropy is to be decomposed consistently. The subsequent claim that the left-hand side becomes larger under heterophily because the joint entropy and H(N_i^+|x_i) 'remain the same' is therefore unsupported. Since this is the only analytical bridge between the δH score and the paper's heterophily criterion, it needs to be either rigorously derived or replaced with a direct empirical characterization.","section":"§3.3, Eq. (12)"},{"comment":"The transition from the Bayesian experimental-design objective to the conditional-entropy estimator in Eq. (11) is asserted rather than derived. Proposition 3.2 and 3.3 present an information-gain utility U(B|ξ) and a Monte Carlo approximation in Eq. (10), but the text does not show how maximizing Eq. (10) reduces to computing pairwise differences of conditional entropies. Moreover, the 'intervention' in §4.1 rescales node features with Gaussian noise; this is not a do-type intervention that sets variables or modifies the data-generating mechanism, so it does not identify causal effects in the Pearl sense. The equivalence claimed between the objective and δH therefore remains unsupported and is central to the paper's causal framing.","section":"§3.2–§4.1, Propositions 3.2–3.3"},{"comment":"The case study on Texas does not validate the causal interpretation. It shows that δH is larger on label-heterophilic edges than on homophilic edges (δ1 = 0.2434 vs δ0 = 0.1534) and that the number of directed edges modified is mostly on heterophilic edges, but this only demonstrates correlation of δH with label disagreement, not that the identified orientation corresponds to true cause-effect relationships. No ground-truth causal graph is available, and the method's structural modifications are not benchmarked against an oracle or against a non-causal asymmetry heuristic. Without such a comparison, the reported gains in Table 4 could be attributed to generic graph augmentation rather than to causal discovery.","section":"§5.3, Case study"}],"minor_comments":[{"comment":"In the abstract, 'the our model' should be 'our model'.","section":"Abstract"},{"comment":"The keyword 'Heterophiliy' is a typo and should be 'Heterophily'.","section":"Keywords"},{"comment":"The notation 'N_I nodes' is not defined; the algorithm should clarify whether this is the node-intervention ratio r_c or an absolute count.","section":"§4.2, Algorithm 1"},{"comment":"The simplification 'O(|V|^2+|E|) = O(N^2)' is misleading because |E| can itself be O(N^2); the statement should read O(|V|^2 + |E|) rather than implying equality to O(N^2).","section":"§4.3, Complexity analysis"},{"comment":"The column header 'Chameron' should be 'Chameleon'.","section":"Table 2"},{"comment":"The sentence 'As first train an unsupervised node embedding network f in CausalMP, we here calculate node dependency metric by Eq.11 on all the edges. We firstly only train the node embedding network f and calculate the node dependency metric by Eq.(11) on all the edges.' is duplicated and should be reduced to a single coherent statement.","section":"§5.3, Case study"},{"comment":"The lists 'the numbers of edges transferred to directed edges are [6,7,8,4,3]' and '[6,7,8,4,1] of them are heterophilic edges' are inconsistent: the first list has five entries and the second list has five as well, but the last entries (3 vs 1) do not match; the intended numbers should be clarified.","section":"§5.3, Case study"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are interesting and the ablation study is reasonably thorough, but the paper's title and framing overstate the causal contribution. The δH estimator in Definition 3.4 is not a valid causal orientation criterion, and Eq. (12) is not a correct derivation. I believe the authors can either provide a rigorous causal-identifiability argument with supporting synthetic experiments, or substantially reframe the contribution as an asymmetric-dependency-based graph augmentation heuristic, removing or heavily qualifying the causal claims. The current gap between the claimed causal discovery and the actual estimator is a load-bearing issue that should be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a useful graph-rewiring heuristic with consistent empirical gains, but the causal-structure story does not survive contact with the math. Treat it as augmentation, not causal discovery.\n\nWhat's new: the specific pipeline — conditional-entropy asymmetry to reorient edges, MI-based edge addition, iterative refinement, and a shared encoder trained on original and modified graphs with a consistency loss — is not in the cited literature. The ablations in Table 4 are the right way to test components, and the Texas case study confirms the score separates heterophilic from homophilic edges. Link prediction results are strong across both regimes, and the few-shot classification improvements are small but consistent. That is a plausible reusable module.\n\nSoft spots: the central estimator is not a causal one. Definition 3.4 takes the absolute value of a conditional-entropy difference, which throws away exactly the sign needed to orient an edge; the text then says the direction is determined by the sign 'inside the absolute value,' which the definition doesn't expose. More fundamentally, conditional-entropy asymmetry is not a causal identifiability criterion without extra structural assumptions; for independent variables with unequal marginal entropies, δH can be large. The step from the Bayesian objective (Eq. 8-10) to δH is asserted, not derived, and Eq. (12) is not a valid entropy identity as written (it appears to assume a conditional independence that is never stated). So 'learns the explicit causal structure' overstates what the algorithm does. The gains may come from generic rewiring rather than cause-effect discovery. The case study validates separation, not direction.\n\nReproducibility: λ1/λ2 are never reported, KDE binning and bandwidth are unspecified, and no code is released. For a method whose behavior depends on those thresholds, that is a real gap.\n\nWho it is for: researchers working on heterophilic GNNs or graph augmentation. Causal inference readers will not be convinced.\n\nRecommendation: do not desk-reject; send to peer review. The empirical module is worth a serious referee. Ask for code, threshold values, and either a full derivation of the causal claim or a reframing as an edge-rewiring heuristic. I would not cite it in its current form, but I would revisit once those are fixed.","headline":"CausalMP is a useful graph-rewiring heuristic with real empirical gains, but the causal-structure claim is not supported by the estimator as written; treat it as augmentation, not causal discovery.","tokens_in":17052,"tokens_out":7107,"would_cite":false,"duration_ms":63275,"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":"CausalMP claims that the asymmetry of conditional entropy between connected nodes reveals causal, heterophilic edges, and rewiring the graph this way improves link prediction and node classification.","keywords":["heterophilic graphs","causal inference","message passing","graph neural networks","link prediction","conditional entropy","graph structure learning","node classification"],"falsifier":"Run CausalMP on a synthetic graph where the ground-truth causal directions and heterophily labels are known. Compute $\\delta_H$ for all edges and check whether it ranks heterophilic edges above homophilic ones and recovers the true direction. A second decisive test: replace $\\delta_H$ with a symmetric score (e.g., $I(X_i;X_j)$) in the identical pipeline; if AUC on heterophilic datasets does not drop, the directional causal claim is unsupported.","tokens_in":16002,"feed_emoji":"🔗","tokens_out":6970,"duration_ms":88546,"temperature":0.7,"pith_summary":"The paper tries to establish that a graph's heterophilic edges—links connecting nodes of different classes—can be detected by the asymmetry of conditional entropy between the two endpoints, and that rewriting the graph according to this causal signal improves GNN link prediction and node classification. The proposed CausalMP pipeline repeatedly intervenes on node features with Gaussian noise, estimates $\\delta_H(i,j)=|H(X_j|X_i)-H(X_i|X_j)|$ for each edge, converts high-asymmetry undirected edges into directed edges, adds edges with high mutual information, and trains two shared encoders on the original and modified graphs with a consistency penalty. If correct, the method gives a task-independent, explicit causal structure of the graph rather than a task-specific denoising heuristic, and it works on both heterophilic and homophilic graphs. The reported experiments show it outperforms link-prediction baselines on nine datasets and that training on the learned structure improves the few-shot node classification accuracy of several base models.","feed_headline":"Entropy asymmetry rewires graphs and beats GNN link predictors","feed_subtitle":"A conditional-entropy test tells which node influences its neighbor, turning heterophilic edges into directed ones.","key_machinery":"The load-bearing object is the cause-effect significance score $\\delta_H(i,j)=|H(X_j|X_i)-H(X_i|X_j)|$, the absolute difference of conditional entropies in the two directions of an edge; a large value flags a prominent cause-effect relationship and a high probability of heterophily. It is estimated by Gaussian-noise interventions on node features, Monte Carlo repetition, and kernel density estimation of the embedding distribution. Around this score, CausalMP builds an iterative loop that directs high-$\\delta_H$ edges, adds high-mutual-information edges, and optimizes a weighted sum of reconstruction losses on the original and modified graphs plus an embedding-consistency penalty.","core_discovery":"The paper's central claim is that intervention-based causal inference can identify heterophilic message-passing: a node pair whose two conditional entropies differ markedly, $\\delta_H(i,j)=|H(X_j|X_i)-H(X_i|X_j)|$ large, is likely to be a causal, heterophilic pair, with the cause being the endpoint that better predicts the other. CausalMP operationalizes this by rescaling node features with Gaussian noise (the intervention), embedding the intervened graphs with an unsupervised GNN, estimating conditional entropies by kernel density estimation, and iteratively pruning and directing edges whose $\\delta_H$ exceeds a threshold while adding edges whose mutual information exceeds another. The resulting causal adjacency matrix $A_c$ is trained jointly with the original graph through two encoders with shared weights, a reconstruction loss on each graph, and a consistency penalty that keeps the embeddings of the two structures aligned. The paper argues this aligns the graph topology with true node dependencies, mitigating feature smoothing over heterophilic neighbors.","pith_inferences":["One implicit testable extension is to replace the entropy-asymmetry score with a symmetric dependence measure (e.g., mutual information alone) in the same pipeline; if link prediction gains persist, the causal direction is not the operative ingredient, which would suggest the benefit comes from topology augmentation rather than cause discovery.","The $\\delta_H$ criterion could serve as a standalone, label-free heterophily detector for graph datasets, complementing homophily-ratio statistics that require ground-truth labels.","The method quantifies cause-effect with conditional entropy of embeddings; in very high-dimensional feature spaces the KDE estimation may become unstable, so a parametric or learned estimator is a natural next step the authors do not explore.","Because the paper frames the learned structure as causal, a strong out-of-distribution test would be to train on one heterophilic graph and transfer the rewiring rule to another; the invariance assumption (Assumption 3.1) predicts this should generalize better than task-specific rewiring."],"forward_implications":["If the entropy-asymmetry criterion is valid, heterophilic edges can be identified without labels, since only the node features and the graph are needed.","The learned causal structure is explicit and task-agnostic, so the same rewired graph can be reused for link prediction, node classification, and potentially other downstream tasks.","Because the method improves few-shot node classification substantially more than full-label settings, it supplies a practical route for label-scarce heterophilic graphs.","The consistency penalty keeps the modified topology close to the original embedding space, so gains are attributable to the rewiring rather than to a collapse of representations.","The approach works on both heterophilic and homophilic graphs, so it is not a heterophily-specific repair but a general message-passing optimizer."],"supporting_citations":[{"why":"Supplies the Bayesian experimental design and Monte Carlo estimator used to turn the intervention-based optimization into a tractable entropy objective.","marker":"[30]"},{"why":"Justifies that Gaussian noise interventions on node features are sufficient to identify causal connections.","marker":"[24]"},{"why":"Provides the result that the lower bound of mutual information between heterophilic node pairs is negatively correlated with the homophily ratio, grounding the entropy-asymmetry rationale.","marker":"[20]"},{"why":"Gives the relation E[P+_i] = 1 - R_h between expected heterophilic connection probability and homophily ratio used in the heterophily insight.","marker":"[13]"},{"why":"Supplies Pearl's causal hierarchy and do-operator framework that the intervention-based formulation is built on.","marker":"[21]"},{"why":"Provides CCA-SSG, the unsupervised embedding network used to produce node embeddings for the dependency estimation.","marker":"[38]"},{"why":"The counterfactual link prediction baseline whose causal approach CausalMP extends and outperforms in the experiments.","marker":"[39]"}],"fun_headline_variants":["Causal entropy test finds heterophilic links in graphs","Intervention-based causal GNN improves link prediction","Asymmetric entropy reveals causal edges for GNNs","Causal messaging rewires heterophilic graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim stands on the idea that the absolute difference between the two conditional entropies of a connected node pair reveals which node causes the other and that this asymmetry marks heterophilic edges; if that signal is not causally informative, the rewiring is just another graph augmentation.","fun_headline_variants_meta":{"raw":{"variants":["Causal entropy test finds heterophilic links in graphs","Intervention-based causal GNN improves link prediction","Asymmetric entropy reveals causal edges for GNNs","Causal messaging rewires heterophilic graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000479,"raw_usage":{"total_tokens":2368,"prompt_tokens":937,"completion_tokens":1431,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1371}},"tokens_in":553,"tokens_out":1431,"duration_ms":10921,"temperature":1.0,"reasoning_tokens":1371,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:50:12.525720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CausalMP on a synthetic graph where the ground-truth causal directions and heterophily labels are known. Compute $\\delta_H$ for all edges and check whether it ranks heterophilic edges above homophilic ones and recovers the true direction. A second decisive test: replace $\\delta_H$ with a symmetric score (e.g., $I(X_i;X_j)$) in the identical pipeline; if AUC on heterophilic datasets does not drop, the directional causal claim is unsupported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian experimental design and Monte Carlo estimator used to turn the intervention-based optimization into a tractable entropy objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies that Gaussian noise interventions on node features are sufficient to identify causal connections."},{"cited_title":"On Local Aggregation in Heterophilic Graphs","cited_arxiv_id":"2106.03213","evidence_quote":"Provides the result that the lower bound of mutual information between heterophilic node pairs is negatively correlated with the homophily ratio, grounding the entropy-asymmetry rationale."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the relation E[P+_i] = 1 - R_h between expected heterophilic connection probability and homophily ratio used in the heterophily insight."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides CCA-SSG, the unsupervised embedding network used to produce node embeddings for the dependency estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The counterfactual link prediction baseline whose causal approach CausalMP extends and outperforms in the experiments."}],"review_version":1}