REVIEW 3 major objections 6 minor 1 cited by
Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases
T0 review · 3 major / 6 minor · reviewed 2026-07-15 · grok-4.5
Pith's one-line read Class imbalance on relational databases can be fixed by gating messages per relation and synthesizing minorities that keep their structural roles.
desk verdict Solid first treatment of class imbalance in RDL: modest but consistent gains on RelBench, clean ablations, no load-bearing math claims that fail. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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).
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- 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.
- §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.
- §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.
minor comments (6)
- 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.
- 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.
- 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.
- §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.
- 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.
- 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.
Circularity Check
No significant circularity: empirical method paper whose claims rest on measured gains, not on self-referential derivations or fitted quantities renamed as predictions.
full rationale
Rel-MOSS is a standard empirical ML paper. Its central claim (average gains of ~2.46% B-Acc / 4.00% G-Mean on 12 RelBench entity-classification tasks) is evaluated against public baselines and ablations (Table 1, Table 2, Appendix D). Proposition 4.1 simply rewrites the standard heterogeneous GNN update (Eq. 3) under a minority-proportion assumption to motivate the gating controller; the algebra does not feed any fitted parameter back into a claimed prediction. Proposition 4.2 and the Rel-Syn construction (Eqs. 11–12) define a heuristic relational signature and interpolate in the joint (embedding, signature) space; the claim that the resulting samples are “faithful” is then checked experimentally (t-SNE, ablations w/o Rel-Syn), not asserted by construction. Hyper-parameters γ, ω, memory size U and learning rate are tuned on validation data in the ordinary way and never presented as first-principles predictions. Self-citations are limited to ordinary prior RDL / imbalance literature and do not supply a uniqueness theorem or load-bearing ansatz that forces the reported numbers. The paper is therefore self-contained against external benchmarks; no derivation step reduces by definition or by self-citation chain to its own inputs.
Assumptions & free parameters
free parameters (4)
- γ (signature-reconstruction weight)
- ω (signature-distance weight)
- memory-bank size U
- interpolation factor λ ~ Beta
assumptions (3)
- domain assumption Standard heterogeneous GNN message-passing (Eq. 3) is a valid inductive bias for RDB entity graphs.
- ad hoc to paper Minority discriminative signal contracts proportionally to the expected minority-neighbor fraction (Prop. 4.1).
- ad hoc to paper Entity-type histograms and fan-in/fan-out counts form a sufficient relational signature for faithful synthesis (Prop. 4.2).
invented entities (2)
-
relation-wise gating controller (Rel-Gate)
-
relation-guided minority synthesizer (Rel-Syn) with relational signatures
Cite this review
Pith. "Pith review of Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases." pith.science (2026). https://pith.science/paper/4ULW65GP
@misc{pith2026260307916,
author = {Pith},
title = {Pith review of: Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases},
year = {2026},
howpublished = {\url{https://pith.science/paper/4ULW65GP}},
note = {Machine review of arXiv:2603.07916}
}
read the original abstract
In recent advances, to enable a fully data-driven learning paradigm on relational databases (RDB), relational deep learning (RDL) is proposed to structure the RDB as a heterogeneous entity graph and adopt the graph neural network (GNN) as the predictive model. However, existing RDL methods neglect the imbalance problem of relational data in RDBs and risk under-representing the minority entities, leading to an unusable model in practice. In this work, we investigate, for the first time, class imbalance problem in RDB entity classification and design the relation-centric minority synthetic over-sampling GNN (Rel-MOSS), in order to fill a critical void in the current literature. Specifically, to mitigate the issue of minority-related information being submerged by majority counterparts, we design the relation-wise gating controller to modulate neighborhood messages from each individual relation type. Based on the relational-gated representations, we further propose the relation-guided minority synthesizer for over-sampling, which integrates the entity relational signatures to maintain relational consistency. Extensive experiments on 12 entity classification datasets provide compelling evidence for the superiority of Rel-MOSS, yielding an average improvement of up to 2.46% and 4.00% in terms of Balanced Accuracy and G-Mean, compared with SOTA RDL methods and classic methods for handling class imbalance.
Forward citations
Cited by 1 Pith paper
-
Incremental Evaluation and Training in Relational Deep Learning
Relational deep learning models can be evaluated and updated incrementally over time, and incremental fine-tuning matches or beats from-scratch retraining on most standard benchmark tasks.
Reviewed July 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.