{"id":"0a445022-58fb-4015-b620-4a2d17d8093f","arxiv_id":"2506.02243","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Selective task-aware attribute promotion into graph nodes improves GNN classification on relational and tabular data compared to schema-based and heuristic graph construction.","lead":"This paper proposes auGraph, a method that improves graph-based machine learning on tabular and relational data by promoting selected attribute values into extra graph nodes, guided by task-relevant scores. It reports accuracy gains over plain schema-based graphs and heuristic baselines on synthetic, hepatitis, and mushroom datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GNN-gain proxy in Section 3.2 is unvalidated and may not be well-defined for a fixed heterogeneous GNN; since Table 1's relational wins use sGNN, this assumption carries the headline claim.","rationale":"Good-faith reading: auGraph's contribution is to make graph construction task-aware by promoting attributes selected through scoring metrics, and the central claim is that this consistently beats schema-based and heuristic graph construction. The strongest empirical support is Table 1: the relational rows use auGraph(top-k, sGNN), while the tabular row uses auGraph(top-k, sMI). The most load-bearing link is therefore the GNN-gain score. The reader's weakest-assumption identification is correct: Section 3.2's fixed-model proxy is never validated against retrained models. I add a concrete technical detail: under the stated HeteroConv/SAGEConv architecture, a graph augmented with a new attribute relation R_A has a new edge type, and a model trained only on G_REG has no trained message-passing parameters for it; the paper does not explain how Eval_{f_theta}(G(A)) is computed. If this forward pass uses randomly initialized or copied weights, the score is essentially evaluating noise. Even if the forward pass is well-defined, 'improving message passing without retraining' is not a general property of GNNs, and no experiment checks it. This matters because losing the sGNN rows would leave the relational evidence for the central claim without support; MI alone carries only the single mushroom tabular result. Other concerns raised by the reader, such as missing error bars, the small number of datasets, and selecting the validation-best metric, are real but secondary relative to the unvalidated proxy. The appropriate verdict remains conditional, with the specific condition that GNN-gain be validated against retraining or replaced by an explicitly trained selector before the relational claims are accepted.","tokens_in":7815,"tokens_out":6855,"duration_ms":68526,"concrete_test":"Run the released auGraph code and inspect the exact forward pass used for sGNN on G(A). First, verify whether a fixed HeteroConv model can evaluate a graph with the unseen relation type R_A without retraining; if the parameters for R_A are randomly initialized or copied from another relation, the score is suspect. Second, on the synthetic dataset, compute GNN-gain rankings for all candidate attributes and compare them against rankings obtained by retraining the GNN from scratch (same hyperparameters) on each augmented graph and evaluating on validation. Report rank correlation and check whether the top-3 attribute sets agree. If the retrained ranking disagrees substantially with GNN-gain, the proxy is invalid and the relational results in Table 1 do not support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 defines sGNN(A) = Eval_{f_theta}(G(A)_REG) - Eval_{f_theta}(G_REG), with f_theta trained once on G_REG and kept fixed. The paper asserts that if inclusion improves message passing, performance should increase even without retraining. Two problems arise. First, under the stated architecture (two-layer heterogeneous GNN using HeteroConv/SAGEConv, Section 4.1), augmenting with attribute A creates a new relation R_A and a new edge type (attribute-value to row). A fixed HeteroConv model normally has no trained message-passing parameters for an unseen relation type; the paper does not specify how Eval_{f_theta}(G(A)) is actually computed. If the forward pass relies on randomly initialized or copied weights, the score measures noise rather than utility. Second, even if the forward pass is well-defined, the proxy is never validated: no experiment compares GNN-gain's attribute ranking with rankings obtained by retraining the GNN on each candidate augmented graph. Because Table 1 reports auGraph(top-k, sGNN) for both relational settings, the headline relational improvements (F1 0.939 vs 0.930 on hepatitis; 0.773 vs 0.734 on synthetic data) depend on this specific unverified assumption. If the proxy misranks attributes, the top-k set can be suboptimal, and the reported relational wins may not reflect a genuine advantage of task-aware augmentation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes auGraph, a graph-construction framework for relational and tabular data. Starting from a relational entity graph (REG), it promotes selected attributes into new attribute-value nodes linked to the tuples that carry those values. Selection is guided by four scoring functions—mutual information, entropy gain, path disagreement, and a GNN-gain proxy—and is performed iteratively up to a budget k or an early-stopping threshold tau. The framework is evaluated on one synthetic relational dataset, the hepatitis multi-table dataset, and the mushroom tabular dataset, with node classification as the target task. The authors report that auGraph outperforms schema-based and heuristic baselines on these datasets.","tokens_in":8075,"tokens_out":6613,"duration_ms":58604,"significance":"The idea of making graph construction task-aware via attribute promotion is well-motivated, and the proposed scoring functions are simple and interpretable. The release of a public implementation supports reproducibility. If the empirical claims hold up, the method could be a useful addition to the relational and tabular GNN toolbox. However, the current evidence is too limited: the GNN-gain score, which drives the relational results, is not fully specified for unseen relation types and is never validated against retraining; the experiments lack error bars and significance tests; and the evaluation covers only three datasets. These gaps are load-bearing for the headline claim of consistent improvement, so the paper requires substantive revision rather than minor polishing.","major_comments":[{"comment":"The GNN-gain score is defined as sGNN(A)=Eval_{f_theta}(G(A)_REG) - Eval_{f_theta}(G_REG) with f_theta trained once on G_REG and kept fixed. This is not well-defined for the architecture described in Section 4.1. Promoting attribute A introduces a new relation R_A and a new edge type between attribute-value nodes and row nodes; a fixed HeteroConv/SAGEConv model has no message-passing parameters for that unseen edge type. The paper does not state how Eval_{f_theta} on G(A)_REG is computed in this situation (e.g., randomly initialized weights, weight sharing, or an extended forward pass). If untrained weights are used, the score reflects random noise rather than attribute utility. Because the relational rows of Table 1 report auGraph(top-k, sGNN), the headline relational gains depend on this unspecified computation. Please specify the exact forward pass and validate the proxy by comparing sGNN attribute rankings with rankings obtained by retraining the GNN on each candidate augmented graph.","section":"Section 3.2 (GNN Gain)"},{"comment":"The auGraph rows in Table 1 are reported as the metric that performs best on the validation set among four scoring functions. This best-of-four selection, together with the lack of standard deviations (Random-k is averaged over three runs only, and no variance is reported for any configuration), makes the 'consistently outperforms' claim hard to evaluate. Selection on the validation set is a legitimate model-selection procedure, but it should be reported as such, with per-metric test results, multiple seeds, and significance tests; otherwise the reported gains could be inflated by selection. Please add these details.","section":"Table 1 / Section 4.2"},{"comment":"The evaluation is too narrow for the paper's general claims. Only one real relational dataset (hepatitis) is used, the relational evidence otherwise comes from a synthetic dataset, and the tabular evidence is a single UCI dataset (mushroom). The abstract and Section 5 claim a unified framework for relational and tabular data; that claim requires a broader benchmark suite (e.g., additional CTU datasets or RelBench) or a more modest statement of scope.","section":"Section 4.1 / 4.2"}],"minor_comments":[{"comment":"The word 'mechanishms' should be 'mechanisms'.","section":"Section 1"},{"comment":"The word 'ealrly-stopping' should be 'early-stopping'.","section":"Section 3.3"},{"comment":"The word 'Singe-table' should be 'Single-table'.","section":"Figure 1 caption"},{"comment":"The normalization constant Z is not defined; please state whether it depends on the attribute A or is the same for all candidates.","section":"Section 3.2 (Path Disagreement)"},{"comment":"The entropy-gain metric is undefined when a d-hop neighborhood contains no training nodes; please specify the convention used in that case.","section":"Section 3.2 (Entropy Gain)"},{"comment":"The claim of being 'the first framework to pose graph construction as a task-aware feature selection problem' is too strong without a more detailed comparison to GFS [19] and other feature-selection-based graph construction methods.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript currently reports only the validation-best auGraph result; I would ask the authors to provide a full per-metric table. The code release is a strength. I did not find evidence of self-citations being used to support the method; the circularity concern identified by the skeptical reader is better framed as an unspecified implementation and validation gap in the GNN-gain proxy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful new thing here is the selection loop: instead of promoting every attribute to a node, as GFS does, auGraph scores candidate attributes and iteratively promotes only the ones that look useful, recomputing scores after each promotion. That framing — graph construction as task-aware feature selection — is sensible, and the paper shows a consistent if narrow pattern: selective promotion beats all-promote and random promotion on the three benchmarks they tried. Credit where due: the code is public, the method is described clearly enough to reimplement, and the authors are honest that the expressivity link is weak.\n\nThe soft spots are real, though. Table 1 has no error bars or significance tests; three datasets is thin for a universal claim. They report the auGraph number as the best of four scoring functions chosen on validation, which inflates test performance relative to the baselines, and they don't show the other scoring functions' test numbers. The GNN-gain proxy in Section 3.2 is the biggest problem: a fixed two-layer heterogeneous GNN trained on the base REG has no trained message-passing weights for a newly added attribute-relation edge type. The paper never says how Eval_{f_theta}(G(A)) is computed — random init? copied weights? — and there's no experiment comparing GNN-gain's ranking against actually retraining. Since the relational wins in Table 1 use sGNN, the headline relational claim rests on this unvalidated proxy. I don't think the central idea is wrong; I just think the current evidence doesn't support the strong statement in the abstract.\n\nThe novelty claim is also a bit overstated. Promotion itself is from GFS [19]; the new bit is selective, iterative choice. That's a legitimate step forward, but I'd frame it as an incremental improvement rather than the 'first task-aware graph construction' claim in the conclusions.\n\nWho gets value: this is a methods paper for people building GNN pipelines over relational or tabular data. It deserves a serious referee; a good reviewer will push for more datasets, error bars, results for all scoring metrics, and a clean explanation/validation of the GNN-gain proxy. I'd like to see the fixed-model proxy checked first, because if it doesn't hold, the relational results get shaky, but the method can still work with the other scores.","headline":"Useful idea—selective task-aware attribute promotion beats all-promote—but the GNN-gain proxy is under-specified and the empirical support is thinner than the abstract claims.","tokens_in":8649,"tokens_out":1888,"would_cite":true,"duration_ms":17124,"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":"auGraph treats graph construction for GNNs as task-aware feature selection and shows that promoting only the useful attributes into nodes beats schema-based and heuristic baselines.","keywords":["graph neural networks","relational databases","tabular data","graph construction","attribute promotion","feature selection","node classification"],"falsifier":"A decisive test: on a relational dataset with known label-relevant attributes, for each attribute compare the GNN-gain proxy — the validation score of a frozen GNN on the graph augmented with that attribute — with the validation score of a freshly retrained GNN on the same augmented graph; if these rankings disagree substantially, the selection procedure rests on an unreliable proxy and the reported gains would not be expected to transfer.","tokens_in":7548,"feed_emoji":"📊","tokens_out":12084,"duration_ms":98647,"temperature":0.7,"pith_summary":"The paper argues that the graph a GNN learns on should be built for the prediction task rather than fixed by the data schema. It introduces auGraph, which starts from the standard relational entity graph and repeatedly promotes non-key attributes into new attribute-value nodes, guided by scores that estimate how much each attribute helps the task. The best auGraph configuration raises F1 from 0.734 to 0.773 on synthetic relational data, from 0.930 to 0.939 on the hepatitis database, and reaches 0.985 on the mushroom table, beating the best heuristic baseline at 0.969. The practical stake is that tabular and relational GNN users get a principled alternative to hand-joining tables or blindly promoting all columns.","feed_headline":"Task-aware graph construction beats fixed schema baselines","feed_subtitle":"Scoring attributes and promoting the useful ones beats schema-based and heuristic graphs on tabular and relational data.","key_machinery":"The carrying mechanism is attribute promotion: a candidate non-key attribute $A$ is lifted into a new unary relation $R_A$ with table $T_A = \\pi_A(T_i)$, so each distinct value becomes a vertex and each value vertex is linked to the row vertices that carry it. Four scoring functions rank candidates: $s_{MI}$ measures mutual information between the attribute and the label on the fully joined training table; $s_{ent}$ measures the drop in label entropy in $d$-hop neighborhoods after augmentation; $s_{dis}$ penalizes paths through the attribute that connect training nodes with different labels; and $s_{GNN}$ compares validation performance of a fixed trained GNN on the augmented versus base graph, without retraining. The iterative loop promotes the top-scoring attribute, recomputes all scores, and stops when the budget $k$ is reached or the best score falls below threshold $\\tau$, so later promotions favor complementary rather than redundant attributes.","core_discovery":"auGraph's central claim is that attribute-level information should enter a relational entity graph selectively, as task-aware structure, rather than staying as flat features or being added wholesale. For a chosen non-key attribute $A$, the framework extracts the unary table $T_A = \\pi_A(T_i)$, adds one node per distinct value, and connects each value node to every row containing that value; it then repeats this promotion with one of four scores — mutual information, entropy gain, path disagreement, or GNN gain — recomputing scores after each step until a budget $k$ or threshold $\\tau$ is reached. The original schema is preserved; only structure is added. In the paper's experiments, the validation-selected auGraph configuration outperforms the relational entity graph, all-promote, random-$k$, and kNN baselines: F1 0.773 versus 0.734 on synthetic relational data, 0.939 versus 0.930 on hepatitis, and 0.985 versus 0.969 on mushroom, with the GNN-gain score best on relational tasks and mutual information best on the tabular task.","pith_inferences":["The fixed-model GNN-gain proxy, if it holds up, implies a cheap recipe for graph construction: train one model on the base graph, score all candidate augmentations with that frozen model, and train only once more on the final graph; a natural check is to compare this proxy against retrained performance on more datasets.","All four scoring functions need labels, so a semi-supervised setting would require pseudo-labels or representation-based scores; the paper does not address this case.","A testable extension suggested by the conclusion is to align augmentation with label agreement under 1-WL colourings; one concrete hypothesis is that label-homogeneous attribute neighborhoods predict downstream accuracy better than the current scores do.","The relational-versus-tabular split in best metric hints that model-based scores matter most where the base graph already carries structure, while statistical scores suffice where the graph is nearly empty; a broader benchmark could test whether that split is stable."],"forward_implications":["On relational data, a GNN can beat the standard schema-derived relational entity graph without changing the GNN architecture, solely by promoting a few validation-selected attributes into nodes.","Promoting every attribute can hurt, so brute-force structural augmentation is not a safe default; task-aware selection is what delivers the gains.","For single-table data, auGraph supplies a graph construction that beats both all-promote and nearest-neighbour graphs, so tabular GNNs need not depend on feature-space geometry.","The best scoring metric differs by regime — GNN gain for relational tasks, mutual information for the tabular task — so choosing among the four scores by validation performance is part of the method rather than a tuning detail.","Because scores are recomputed after every promotion, the construction favors compact, complementary graphs, which matters when the base graph is large."],"supporting_citations":[{"why":"Defines the relational entity graph that auGraph takes as its base structure and as the primary relational baseline.","marker":"[6]"},{"why":"Supplies the all-promote graph construction baseline, which promotes every attribute and is outperformed by auGraph.","marker":"[19]"},{"why":"Supplies the k-nearest-neighbour graph baseline used in the tabular experiments.","marker":"[5]"},{"why":"Source of the hepatitis relational dataset used for the multi-table experiments.","marker":"[13]"},{"why":"Source of the mushroom tabular dataset used for the single-table experiments.","marker":"[17]"}],"fun_headline_variants":["Task-aware attribute promotion beats schema graphs","auGraph: promote useful attributes to nodes","Selective attribute-to-node promotion boosts GNNs","Task-aware graph construction outperforms fixed schemas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a fixed model's validation score on an augmented graph, without retraining, ranks attributes in roughly the same order as a freshly retrained model's score would.","fun_headline_variants_meta":{"raw":{"variants":["Task-aware attribute promotion beats schema graphs","auGraph: promote useful attributes to nodes","Selective attribute-to-node promotion boosts GNNs","Task-aware graph construction outperforms fixed schemas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1289,"prompt_tokens":960,"completion_tokens":329,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":284}},"tokens_in":576,"tokens_out":329,"duration_ms":3714,"temperature":1.0,"reasoning_tokens":284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:27:03.603219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: on a relational dataset with known label-relevant attributes, for each attribute compare the GNN-gain proxy — the validation score of a frozen GNN on the graph augmented with that attribute — with the validation score of a freshly retrained GNN on the same augmented graph; if these rankings disagree substantially, the selection procedure rests on an unreliable proxy and the reported gains would not be expected to transfer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the relational entity graph that auGraph takes as its base structure and as the primary relational baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the k-nearest-neighbour graph baseline used in the tabular experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the mushroom tabular dataset used for the single-table experiments."}],"review_version":1}