{"id":"536d9c8e-8186-410f-89f1-6f0153ce6986","arxiv_id":"2507.12562","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Rel-HNN models each tuple as a hyperedge over attribute-value nodes and reports large accuracy gains, but its hypergraph construction appears to include the target label as an input node.","lead":"Rel-HNN turns relational database rows into hypergraph edges and attribute-value pairs into nodes, then trains a hypergraph neural network for prediction tasks. The paper claims large accuracy gains over prior graph methods and multi-GPU speedups, but the central setup may leak the target label into the model's input.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 builds hypergraph nodes from every column, including the target label column, so each training hyperedge contains the true label as an input node; unless an exclusion exists that the paper never states, the reported AUROC/RMSE gains are label leakage.","rationale":"The paper's central claim is that rel-HNN significantly outperforms state-of-the-art methods on classification and regression over relational databases, with the empirical tables as the only support. For that claim to hold, the evaluation must measure generalization rather than the model reading the target label from its input. Algorithm 1, as written, makes the label an attribute-value node in every training hyperedge, and Equations 2, 4, and 5 route that node's embedding into the final prediction embedding. This is a structural circularity, distinct from tuning choices or baseline comparisons: it invalidates the reported accuracy evidence unless the implementation silently excludes the label column, in which case the paper still fails to document the exclusion. The reader's weakest_assumption identifies the same issue, and the proposed ablation would settle it directly. The split-parallel speedup results (Figures 5–7) may be independently valid engineering contributions, but they do not rescue the central accuracy claim. No new concern beyond the reader's is needed to justify the REJECT verdict, so the verdict should remain unchanged.","tokens_in":20415,"tokens_out":3904,"duration_ms":43726,"concrete_test":"","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 1, the hypergraph construction loops over all tables, all rows, and all columns (lines 3–9), then builds each tuple hyperedge from those same attribute-value nodes (lines 10–16). For the target table T_tg, the label column (e.g., is_fraud) is just another column, so for every training row i the node (label, y_i) is created and included in hyperedge e_i. Equation 2 then forms the initial tuple embedding F^0_e by an MLP over the sum of all node embeddings in e, including the label node; Equations 4 and 5 propagate that signal through all L layers, and the paper states that the final Z^L_e for target-table hyperedges is the predicted class probability. Thus the target value is an input feature, not merely a supervision signal. The paper does not state that the label column is excluded from Algorithm 1 or from the feature encodings in Section 4.1, and the experimental section does not mention any special handling of the target column. If the released implementation does exclude it, the manuscript omits a crucial detail; if it does not, the large AUROC/RMSE improvements in Tables 3 and 5 are explained by label leakage rather than by the hypergraph representation. This is the most load-bearing concern because the central claim of outperforming state-of-the-art methods rests entirely on those tables.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes rel-HNN, a hypergraph neural network for supervised learning on relational databases. The method builds a hypergraph in which each unique attribute-value pair is a node and each tuple is a hyperedge, and it learns embeddings at attribute-value, tuple, and table levels via alternating node-to-hyperedge and hyperedge-to-node message passing. A split-parallel multi-GPU training algorithm is introduced for scalability. The empirical sections report AUROC and RMSE comparisons on nine classification and four regression datasets, as well as training-time speedups. The central claim is that rel-HNN significantly outperforms state-of-the-art graph- and hypergraph-based methods on both tasks.","tokens_in":20656,"tokens_out":7804,"duration_ms":82409,"significance":"If the reported results were valid, the paper would make a useful contribution: the attribute-value hypergraph representation is schema-agnostic, avoids flattening, and the split-parallel algorithm addresses a real scalability need for hypergraph neural networks. The paper also provides a code repository link and reports detailed dataset statistics, which are strengths. However, the evaluation is undermined by a fundamental design issue: Algorithm 1 appears to place the target label column into the input hypergraph, so each training hyperedge contains the true label as one of its nodes. As written, the accuracy and RMSE improvements in Tables 3 and 5 are explainable by label leakage rather than by the proposed architecture. A major revision that excludes the label column from hypergraph construction and re-runs the experiments would be needed before the central claim can be assessed.","major_comments":[{"comment":"Algorithm 1 creates a node for every unique attribute-value pair in every column of every table, including the target table's label column. For each training row i of the target table, the node (label, y_i) is therefore included in hyperedge e_i. Equation (2) forms the initial hyperedge embedding as an MLP over the sum of the node embeddings in e_i, and Equations (4) and (5) propagate this signal through all layers; the text then states that Z^L_e for target-table hyperedges is the predicted class probability. Thus the true label is an input feature, not merely a supervision signal. The manuscript never states that the label column is excluded from Algorithm 1 or from the feature encodings in Section 4.1, and Section 5 reports no experiment that removes the label column. Unless the implementation silently omits the target column, the large AUROC improvements in Table 3 and the 90%+ RMSE reductions in Table 5 are explained by label leakage rather than by the proposed representation. This is the load-bearing support for the paper's central claim.","section":"4.1, Algorithm 1; Equations (2)-(5); Tables 3 and 5"},{"comment":"The task is defined as predicting labels for rows of the target table where the label is unknown, but Algorithm 1 builds a hyperedge for every row from all of its columns. For test rows, either the label values are present in the input, which is leakage, or they are absent, in which case test hyperedges have a different set of nodes than training hyperedges and the model's input distribution changes. The manuscript does not describe how missing target labels are handled during hypergraph construction or inference, so the reported train/test protocol is ambiguous and the experimental comparison cannot be interpreted as a standard supervised-learning evaluation.","section":"Section 2 and Algorithm 1"}],"minor_comments":[{"comment":"The update step omits MLP^0_V and MLP^0_E from the parameter sets, even though these MLPs are used in lines 5-8 and are included in the initialization in line 2; if this is a typo, correct it, and if it is intentional, explain why the first-layer projections are not trained.","section":"Algorithm 2, line 15"},{"comment":"The entries for SameGen, st_loan, and Mutag are concatenated into '0.50300.94110.8812'; fix the formatting so each value is separately readable.","section":"Table 3, ATJ-net row"},{"comment":"The text refers to an undefined 'Equation??', and Equations (7) and (9) sum from i=0 to N even though there are N GPUs; the index range should presumably be i=1 to N.","section":"Section 4.3, Equations (6)-(9)"},{"comment":"Reference [9] is a duplicate of reference [8]; the duplicated citation should be removed or replaced.","section":"References"},{"comment":"The statement that the embedding length of all nodes and hyperedges is 'fixed at two' is unexpectedly small; clarify whether this is literally 2 or a shorthand for a larger dimension, since such an embedding size would likely be too small for the reported tasks.","section":"Section 5.1, experimental settings"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is label leakage in the input representation. I recommend asking the authors to modify Algorithm 1 to explicitly exclude the target column, re-run all classification and regression experiments, and either report updated tables or justify why the current tables remain valid. The split-parallel speedup results are less affected by this issue, but they should also be re-checked after the fix. If the authors cannot provide an implementation without leakage, rejection would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one with leakage in mind. The central empirical claim is invalid as written. Algorithm 1 creates a node for every attribute-value pair across all columns, including the target label column. Each tuple hyperedge then contains its true label node, and the final hyperedge embedding fed to the prediction head is a learned function of that node. For test rows, the construction is even worse: you cannot build the test hyperedge without knowing the label. The paper never states that the label column is excluded, and the experiments give no hint of special handling. So the large AUROC/RMSE gaps in Tables 3 and 5 are most plausibly explained by label leakage, not by the hypergraph representation.\n\nThat said, the paper is not without merit. The attribute-value hypergraph representation is a genuine departure from ATJ-Net, which only uses joinable attributes as vertices, and the two-phase node-to-hyperedge/hyperedge-to-node message passing is a reasonable design on that structure. The split-parallel multi-GPU algorithm—partitioning nodes and exchanging partial hyperedge sums—is a real engineering contribution, and the speedup trends in Figure 5 are internally consistent. The dataset breadth is also nice: nine classification and four regression benchmarks is more than most papers in this area.\n\nThe soft spots beyond the leakage: the “first hypergraph-based representation” claim conflicts with their own citation of ATJ-Net; hyperparameters (embedding dimension 2, L=2) are fixed without ablations; and the experimental setup is under-specified about how cross-validation interacts with hypergraph construction. These are secondary. The leakage is the load-bearing flaw.\n\nThis paper is for someone studying evaluation pitfalls in relational or hypergraph learning, and the split-parallel idea could be reused. But in current form the evaluation needs to be rebuilt from scratch. A referee should reject it, while pointing out that excluding the target column and re-running would be the natural fix. I would not cite the empirical results, though I might cite the parallel algorithm if it were properly evaluated.\n\nRecommendation: send to peer review because the topic and representation are serioous enough to warrant careful referee attention, but expect a clear reject on the current evidence.","headline":"The leakage in Algorithm 1 is real: the label column is built into the hypergraph, so the reported accuracy gains are not trustworthy; the split-parallel training idea may still be salvageable.","tokens_in":21232,"tokens_out":3046,"would_cite":false,"duration_ms":36919,"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":"A hypergraph that makes one node per column-value pair and one hyperedge per tuple outperforms graph baselines on relational learning benchmarks.","keywords":["hypergraph neural network","relational database learning","attribute-value hypergraph","multi-GPU training","table embedding","relational classification and regression","split-parallel training","schema-agnostic representation"],"falsifier":"Run the released code with the target column removed from the attribute-value node construction, or inspect the code to see whether it is already excluded; if the AUROC and RMSE gaps against ATJ-Net vanish or shrink substantially, the empirical advantage is leakage rather than representation.","tokens_in":20168,"feed_emoji":"📊","tokens_out":5645,"duration_ms":59451,"temperature":0.7,"pith_summary":"The paper proposes rel-HNN, a hypergraph neural network for relational databases that creates one node per unique attribute-value pair and one hyperedge per tuple. The model alternates message passing between nodes and hyperedges, with optional learnable per-table embeddings, and is trained end-to-end to predict a target column. On nine classification datasets it reports higher AUROC than GCN, GAT, SPARE, and ATJ-Net on eight, and on four regression datasets it reports lower RMSE across the board. A node-partitioned multi-GPU variant reduces per-epoch training time by up to 3.18x on relational data and 2.94x on benchmark hypergraphs. The paper's claim is that attribute-level hypergraph structure, not schema engineering, is what drives these gains.","feed_headline":"Attribute-value hypergraph beats graph baselines on relational tasks","feed_subtitle":"Reporting top AUROC and lower RMSE than graph baselines on relational benchmarks.","key_machinery":"The load-bearing object is the attribute-value hypergraph: for every table, every row, and every column, a node is created for the pair (column name, cell value), and each row then becomes a hyperedge connecting the nodes of its cells. This turns the schema into an unlabeled incidence structure, so no primary-key/foreign-key extraction is needed and shared values across tables become shared nodes. On this hypergraph, rel-HNN runs two-phase message passing: hyperedge embeddings are sums of node embeddings passed through an MLP, node embeddings are sums of incident hyperedge embeddings passed through another MLP, and optional per-table embedding vectors are concatenated to the hyperedge embeddings to inject global context. The split-parallel algorithm factorizes the hyperedge MLP into a linear part computed locally per GPU plus a nonlinear activation applied after cross-GPU summation of partial hyperedge sums.","core_discovery":"The central claim is that representing each attribute-value pair as a node and each tuple as a hyperedge lets a hypergraph neural network learn richer representations than tuple-as-node graphs, because intra-tuple associations are preserved as hyperedges rather than flattened into pairwise edges. Rel-HNN learns embeddings at three levels: attribute-value nodes, tuple hyperedges, and table embeddings, and uses two-phase message passing that first aggregates nodes into hyperedges and then aggregates hyperedges back into nodes. The paper further claims that splitting the node set across GPUs and exchanging partial hyperedge sums yields near-lossless speedups. Empirically it reports state-of-the-art AUROC on eight of nine classification datasets and lower RMSE on all four regression datasets, with the largest gains on datasets with many tables and columns.","pith_inferences":["If the target column is included in Algorithm 1's node construction, then each training row's hyperedge contains a node encoding the true label, and the final hyperedge embedding used for prediction is a function of that node; the reported accuracy gains could then be explained by label leakage rather than by the representation.","A clean test would rerun the experiments with the target column excluded from node construction; the speedup results would be unaffected, but the AUROC and RMSE comparisons would likely change substantially.","The split-parallel scheduler is orthogonal to the hypergraph representation and could be applied to other hypergraph neural networks, making the speedup claim the more robust contribution if the accuracy claim is confounded."],"forward_implications":["Because the representation treats every column uniformly, the method can be applied to a database without extracting primary-key/foreign-key constraints or doing manual feature engineering.","Hyperedge-level message passing captures co-occurrence of attribute-value pairs inside a tuple, which tuple-as-node graph models do not represent directly.","Explicit per-table embeddings add global context and appear to help most on schemas with many tables; on shallow schemas the non-table variants are competitive.","The split-parallel training reduces per-epoch time on large hypergraphs, while communication overhead can erase the benefit on small datasets."],"supporting_citations":[{"why":"The prior hypergraph-based relational learning method that rel-HNN directly compares against and claims to exceed.","marker":"[4]"},{"why":"The single-pass DAG baseline whose GCN and GAT variants are used as principal competitors.","marker":"[16]"},{"why":"The graph convolution baseline applied to primary-key/foreign-key tuple graphs.","marker":"[21]"},{"why":"The attention-based graph baseline applied to the same tuple graphs.","marker":"[31]"},{"why":"Supplies most of the classification and regression benchmark datasets used in the evaluation.","marker":"[24]"},{"why":"Supplies the two Formula 1 benchmark datasets used for the largest classification tests.","marker":"[26]"},{"why":"The full-graph multi-GPU GNN training approach that motivates the split-parallel design and comparison.","marker":"[32]"}],"fun_headline_variants":["Hypergraph with attribute-value nodes outperforms graphs on RDBs","Split-parallel training gives 3.18x speedup for relational learning","Rel-HNN models tuples as hyperedges, boosting relational accuracy","Fine-grained hypergraph learning achieves top results on RDBs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Algorithm 1 builds hypergraph nodes from every column of every table, including the target label column, so during training each row's hyperedge contains a node that encodes the true label and that node's embedding can be read off to make the prediction.","fun_headline_variants_meta":{"raw":{"variants":["Hypergraph with attribute-value nodes outperforms graphs on RDBs","Split-parallel training gives 3.18x speedup for relational learning","Rel-HNN models tuples as hyperedges, boosting relational accuracy","Fine-grained hypergraph learning achieves top results on RDBs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000495,"raw_usage":{"total_tokens":2416,"prompt_tokens":920,"completion_tokens":1496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1421}},"tokens_in":536,"tokens_out":1496,"duration_ms":16280,"temperature":1.0,"reasoning_tokens":1421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:44:19.499645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code with the target column removed from the attribute-value node construction, or inspect the code to see whether it is already excluded; if the AUROC and RMSE gaps against ATJ-Net vanish or shrink substantially, the empirical advantage is leakage rather than representation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior hypergraph-based relational learning method that rel-HNN directly compares against and claims to exceed."},{"cited_title":"SPARE: A Single-Pass Neural Model for Relational Databases","cited_arxiv_id":"2310.13581","evidence_quote":"The single-pass DAG baseline whose GCN and GAT variants are used as principal competitors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The graph convolution baseline applied to primary-key/foreign-key tuple graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The attention-based graph baseline applied to the same tuple graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The full-graph multi-GPU GNN training approach that motivates the split-parallel design and comparison."}],"review_version":1}