{"id":"bbd4107d-603b-420d-aa8f-cca62d96947e","arxiv_id":"2603.07916","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Rel-MOSS uses per-relation gating plus relational-signature-conditioned oversampling to improve minority-class entity classification on heterogeneous graphs derived from relational databases.","lead":"Rel-MOSS adds a relation-wise gate and a signature-guided synthesizer so GNNs on relational-database graphs stop drowning minority classes. The method lifts balanced accuracy and G-mean on twelve RelBench entity-classification tasks relative to prior RDL and imbalance baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The strongest claim is modest, consistent empirical superiority on a public benchmark, not a theorem that requires the signature to be a true sufficient statistic. The reader's weakest-assumption note correctly flags a heuristic design choice, but that choice is already stress-tested by the paper's own ablations and visualizations (Figure 5): removing Rel-Syn collapses performance on severely imbalanced sets, while the full model produces synthetic points that visually align with the true minority manifold. Because the claim is empirical and the supporting evidence is multi-dataset, multi-metric, and multi-backbone, the heuristic nature of the signature does not threaten the reported improvements. No stronger load-bearing flaw (e.g., data leakage, metric misdefinition, or non-reproducible gains) is apparent from the manuscript. Therefore the ACCEPT verdict with high confidence stands; the only useful next check is tighter statistical characterization of the deltas already shown in Table 1.","tokens_in":22755,"tokens_out":493,"duration_ms":4431,"concrete_test":"Re-run the full Table 1 protocol with a fixed random seed suite (e.g., 10 seeds) and report paired Wilcoxon or bootstrap CIs on the per-dataset B-Acc/G-Mean deltas versus the best baseline; if the mean improvement remains positive and the lower CI bound stays above zero on the majority of the 12 datasets, the headline claim is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is an empirical one: Rel-MOSS improves Balanced Accuracy and G-Mean on 12 RelBench entity-classification tasks relative to RDL and classic imbalance baselines (Abstract, Table 1). The reader's weakest assumption (that the hand-crafted relational signature is a sufficient statistic for linear interpolation to stay on the minority manifold) is a modeling choice, not a load-bearing correctness condition. Proposition 4.2 is informal and the signature is heuristic, yet the claim does not rest on its theoretical sufficiency; it rests on measured gains that survive ablations (w/o Rel-Gate / w/o Rel-Syn), multiple backbones (Appendix D), and ranking metrics (Table 2). No internal inconsistency, circular derivation, or uncheckable mathematical step undermines the reported numbers. Code and public data further reduce verification risk.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper addresses class imbalance in relational deep learning (RDL) for entity classification on relational databases. It models RDBs as heterogeneous entity graphs and proposes Rel-MOSS, consisting of (i) Rel-Gate, a relation-wise gating controller that modulates neighborhood messages by estimated minority-leaning likelihood (Eqs. 9–10), and (ii) Rel-Syn, a relation-guided minority synthesizer that interpolates both gated embeddings and hand-crafted relational signatures (entity-type histograms and fan-in/fan-out counts; Eqs. 11–12) to preserve relational consistency. Optimization combines BCE classification with MSE signature reconstruction (Eqs. 13–16). Empirically, on 12 RelBench entity-classification tasks, Rel-MOSS improves Balanced Accuracy and G-Mean over native RDL variants and classic imbalance methods (SMOTE, GraphSMOTE, GraphSHA, focal loss, ReVar), with ablations, t-SNE/centroid-distance checks, ranking metrics, runtime analysis, and multi-backbone results.","tokens_in":23071,"tokens_out":1290,"duration_ms":10182,"significance":"Class imbalance is practically important for RDB tasks such as fraud and churn, and prior RDL work largely ignored it. Rel-MOSS is a clear first step that is relation-aware rather than a direct transplant of homogeneous-graph oversampling. Strengths include a broad RelBench evaluation (12 datasets), multiple baselines and ablations (Table 1), qualitative evidence that Rel-Gate increases class separation and Rel-Syn produces more faithful minorities (Figs. 4–5), ranking metrics (Table 2), runtime comparison (Table 3), multi-backbone results (Appendix D), and public code. Propositions 4.1–4.2 are informal analyses rather than theorems, and the relational signature is heuristic, but the central claim is empirical and is supported by the reported experiments. If the gains hold under independent re-runs, the work is a useful, deployable contribution to imbalanced RDL.","major_comments":[{"comment":"Table 1 and Abstract: the headline “average improvement of up to 2.46% / 4.00%” is not fully transparent. Per-dataset improvements vary widely (e.g., large gains on f1-driver-top3 and avito-user-clicks; near-zero or slightly negative G-Mean on trial-study-outcome). Please state explicitly how the average is computed (macro over all 12, only over severely imbalanced sets, vs best baseline only, etc.) and report both mean and median (or a full distribution) so the claim cannot be read as uniform superiority.","section":null},{"comment":"§4.3 and Proposition 4.2: the relational signature (1-/2-hop type histograms + fan-in/fan-out) is presented as the mechanism that “maintains relational consistency,” yet it is a fixed hand-crafted statistic with no ablation of its components and no quantitative measure of how well synthetic samples match true minority structural roles beyond t-SNE. Because Rel-Syn is argued to be the main driver (RQ3), a short ablation (signature off vs type-hist only vs full) or a structural fidelity metric would make the central design choice load-bearing rather than purely heuristic.","section":null},{"comment":"§5.1 / Table 1 baselines: GraphSMOTE and GraphSHA are homogeneous-graph methods; the paper states they are “extended” to heterogeneous graphs but does not specify how multi-relation edges and type-specific neighborhoods are handled in the edge predictor / selective mixup. Without that protocol (or a citation to a standard heterogeneous adaptation), it is hard to judge whether the comparison is fully fair. A brief appendix paragraph would resolve this.","section":null}],"minor_comments":[{"comment":"Proposition 4.1 (Eqs. 4–8): the contraction argument drops the nonlinearity under a local Lipschitz assumption and treats π_e,r as fixed; a one-sentence caveat that this is an informal analysis, not a formal proof, would avoid over-reading.","section":null},{"comment":"Figure 3 / Eq. 9: the Softmax is written over Q(X_e)K(H_e,r)^⊤ / √d; clarify whether this is a single-vector attention (scalar gate) or multi-head, and whether H_e,r is mean-pooled before the attention.","section":null},{"comment":"Table 1: several native RDL rows report exact 0.5000 / 0.0000 with no std; state whether these are collapsed runs (all-majority predictions) and how many seeds were used for the non-degenerate entries.","section":null},{"comment":"§5.6 / Table 3: runtime is reported for four datasets only; a note on whether majority down-sampling was used only for timing or also for the main accuracy numbers would avoid confusion.","section":null},{"comment":"Typos / polish: “Descrip�on” in Fig. 2, “predic�ve”, “Ra�ng”; “consist of” → “consists of” (§4.4); “the class imbalance problem of entity classification on RDBs” (Abstract) is slightly awkward.","section":null},{"comment":"Related work: a short pointer to other heterogeneous or multi-relational imbalance methods (if any) beyond homogeneous GraphSMOTE/GraphSHA would strengthen the “first time on RDB entity graphs” claim.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The empirical package is solid for a methods paper and the problem is timely. The main risk is overstated averaging language and under-specified heterogeneous adaptations of the graph-imbalance baselines; both are fixable in revision. I would not block on the informal propositions. Fit for a solid ML venue is good if the averaging and baseline-protocol points are clarified."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is the first paper that actually treats class imbalance as a first-class problem inside the RDL pipeline on RelBench. That is the real contribution. Everything else is engineering around that gap.\n\nWhat they ship is two modules. Rel-Gate is a relation-specific soft gate (cross-attention style) that tries to keep minority signal from being washed out by majority neighbors. Rel-Syn is GraphSMOTE-style interpolation that also mixes a hand-crafted relational signature (1-/2-hop type histograms + fan-in/fan-out counts) so the synthetic nodes stay roughly consistent with the local schema. Neither piece is conceptually deep, but the combination is new for heterogeneous entity graphs coming from RDBs, and they show it works.\n\nThe empirical package is the strong part. Twelve RelBench entity-classification tasks, proper imbalance baselines (SMOTE, GraphSMOTE, GraphSHA, focal, ReVar), ablations of each module, t-SNE and centroid-distance checks, hyper-parameter sweeps, runtime numbers, and even a couple of alternative backbones (HGT, RelGNN). Gains are modest—roughly 2.5 % B-Acc / 4 % G-Mean on average—but they are consistent and largest exactly where native RDL collapses to majority-class predictions. Code is promised; data is public. That is enough to check the claim.\n\nSoft spots are real but secondary. Propositions 4.1–4.2 are informal sketches, not proofs; the signature is a heuristic, not a proven sufficient statistic. On a couple of near-balanced datasets the lift is tiny or slightly negative on one metric. Hyper-parameters (γ, ω, memory size) still need the usual sensitivity tables and significance tests. None of this undercuts the main empirical result.\n\nThis is for people already working on RDL or imbalanced heterogeneous graphs who need a practical fix that does not break the relational structure. It is not a theory paper and does not claim to be. I would send it to referees; the problem is real, the evaluation is thorough enough, and the result is checkable. Worth engaging if you care about making RDL usable on the long-tail tasks that actually matter in production databases.","headline":"Solid first treatment of class imbalance in RDL: modest but consistent gains on RelBench, clean ablations, no load-bearing math claims that fail.","tokens_in":23591,"tokens_out":558,"would_cite":true,"duration_ms":5755,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Class imbalance on relational databases can be fixed by gating messages per relation and synthesizing minorities that keep their structural roles.","keywords":["relational deep learning","class imbalance","heterogeneous entity graph","minority oversampling","relation-wise gating","relational signature","entity classification","RelBench"],"falsifier":"If synthetic minorities generated by Rel-Syn systematically fall outside the true minority structural distribution (visible as large signature or embedding divergence on a held-out test set) while still being labeled minority, balanced-accuracy gains should disappear.","tokens_in":23705,"feed_emoji":"⚖️","tokens_out":548,"duration_ms":5076,"temperature":0.7,"pith_summary":"Relational deep learning turns multi-table databases into heterogeneous graphs so that graph neural networks can classify entities end-to-end. Real databases are heavily skewed: fraudsters, churners, or rare outcomes are dwarfed by ordinary records, and ordinary message-passing lets majority neighbors drown out the minority signal until the two classes become indistinguishable. Rel-MOSS counters this with two relation-aware modules. A per-relation gating controller first estimates how much each neighborhood message leans toward the minority class and then amplifies or suppresses it accordingly. A guided synthesizer next creates new minority samples by interpolating both the gated embeddings and explicit relational signatures (neighbor-type histograms and fan-in/fan-out counts), keeping the synthetic entities inside the true minority structural manifold. On twelve RelBench classification tasks the method improves balanced accuracy by up to 2.46 percent and G-Mean by up to 4 percent over both standard relational models and classic imbalance remedies, while remaining competitive on balanced data.","feed_headline":"Relational databases need gated messages to stop majority drowning minorities","feed_subtitle":"Per-relation gates plus structure-preserving synthesis lift balanced accuracy on twelve real RDB tasks","key_machinery":"Rel-Gate (a relation-wise attention-style controller that estimates minority-leaning likelihood of each relation’s messages and re-weights them) together with Rel-Syn (interpolation of both embeddings and hand-crafted relational signatures that preserves structural role).","core_discovery":"The first systematic treatment of class imbalance for entity classification on relational databases shows that relation-wise gating of neighborhood messages plus signature-guided minority oversampling restores distinguishability between rare and common entities and yields measurable gains in balanced accuracy and G-Mean across twelve real-world tasks.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Rel-MOSS gates relations to revive minority entities in RDB graphs","Relation-wise gates and synthesizers fix class imbalance in RDL","Gated messages plus synthetic minorities balance RDB entity classes","First RDB imbalance treatment: relation gates and minority synthesis","Relation-gated oversampling restores rare entities in databases"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that a simple hand-crafted signature of neighbor-type counts and relation fan-in/fan-out fully captures an entity’s structural role, so linear interpolation of those signatures stays inside the true minority manifold.","fun_headline_variants_meta":{"raw":{"variants":["Rel-MOSS gates relations to revive minority entities in RDB graphs","Relation-wise gates and synthesizers fix class imbalance in RDL","Gated messages plus synthetic minorities balance RDB entity classes","First RDB imbalance treatment: relation gates and minority synthesis","Relation-gated oversampling restores rare entities in databases"]},"model":"grok-4.5","effort":"low","cost_usd":0.003694,"raw_usage":{"total_tokens":1193,"prompt_tokens":770,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":36940000,"prompt_tokens_details":{"text_tokens":770,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":336,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":770,"tokens_out":87,"duration_ms":2923,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T12:56:27.054886+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"If synthetic minorities generated by Rel-Syn systematically fall outside the true minority structural distribution (visible as large signature or embedding divergence on a held-out test set) while still being labeled minority, balanced-accuracy gains should disappear.","supporting_citations":[],"review_version":1}