Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases

Jun Yin; Peng Huo; Bangguo Zhu; Hao Yan; Senzhang Wang; Shirui Pan; Chengqi Zhang

cs.AIMar 9, 2026

Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases

Authors: Jun Yin, Peng Huo, Bangguo Zhu, Hao Yan, Senzhang Wang, Shirui Pan, Chengqi Zhang

Organizations: Department of Data Science and Artificial Intelligence, Hong Kong Polytechnic University, Hong Kong SAR, China · National Super Computing Center, Tianjin, China · School of Computer Science and Engineering, Central South University, Hunan, China · School of Information and Communication Technology, Griffith University, Queensland, Australia

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.

Figures & tables

Figure 1: Illustration of a) . Aggregation with equal importance leads to indistinguishable representations and b) Decisive information resides in local relational structures.

Figure 2: Illustrative example of a) . Relational database schema, and b) . Tables within the relational database.

Figure 3: Overview of Rel-MOSS. Beginning with the entity feature encoding , the original entity features with diverse modalities is converted into unified representations by modality-specific feature encoder. For entity $e$ , the relation-wise gating controller modulates the neighborhood message from each relation according to the gating factors $\{\Psi_{e,r}\}$ . Subsequently, based on the gated representations, the relation-guided minority synthesizer integrates entity relational signatures into over-sampling process. Finally, the optimization of Rel-MOSS consists of the entity classification and relational signature reconstruction.

Dataset	f1-driver-top3		f1-driver-dnf		avito-user-clicks		avito-user-visits
Metric	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean
RDL	0.5000	0.0000	0.6274 ±0.0692	0.5835 ±0.0995	0.5000	0.0000	0.5000	0.0000
RDL-HGT	0.5124 ±0.0175	0.1179 ±0.1670	0.5497 ±0.0324	0.3647 ±0.1116	0.5000	0.0000	0.5000	0.0000
RelGNN	0.4976 ±0.0034	0.0927 ±0.1311	0.5272 ±0.0066	0.2973 ±0.0332	0.5000	0.0000	0.5000	0.0000
RDL-Focal	0.5000	0.0000	0.5871 ±0.0094	0.4940 ±0.0485	0.5491 ±0.0080	0.3465 ±0.0294	0.6086 ±0.0098	0.5376 ±0.0265
SMOTE	0.7509 ±0.0307	0.7349 ±0.0242	0.6304 ±0.0321	0.5857 ±0.0636	0.5984 ±0.0142	0.5105 ±0.0415	0.6255 ±0.0063	0.5947 ±0.0216

Table 1: Evaluation results of Rel-MOSS and baselines on 12 RDB entity classification datasets, in the form of mean

{\pm}

std . For both Balanced Accuracy (B-Acc) and G-Mean , higher values indicate better performance. Underline marks the best baseline, and bold indicates that Rel-MOSS or Rel-MOSS variants outperform the best baseline. Improvement is defined as

(\mathtt{Rel\text{-}MOSS}-\mathtt{Best\text{-}Baseline})/\mathtt{Best\text{-}Baseline}

Figure 4: Comparison between Rel-MOSS and w/o Rel-Gate in terms of a). Euclidean distance, and b). Manhattan distance.

Figure 5: Visualization of entity representations learned by SMOTE, GraphSMOTE, and Rel-MOSS.

Dataset	f1-driver-top3		f1-driver-dnf		event-user-repeat		event-user-ignore
Metric	PR-AUC	ROC-AUC	PR-AUC	ROC-AUC	PR-AUC	ROC-AUC	PR-AUC	ROC-AUC
RDL	0.2796 ±0.0035	0.7572 ±0.0112	0.2030 ±0.0044	0.7091 ±0.0094	0.7012 ±0.0103	0.7674 ±0.0102	0.4975 ±0.0087	0.8092 ±0.0088
RelGNN	0.3085 ±0.0149	0.7792 ±0.0105	0.2081 ±0.0100	0.6911 ±0.0123	0.6323 ±0.0106	0.7212 ±0.0145	0.4525 ±0.0069	0.8301 ±0.0075
SMOTE	0.2993 ±0.0250	0.7369 ±0.0185	0.1996 ±0.0057	0.7017 ±0.0087	0.7155 ±0.0099	0.7808 ±0.0091	0.5028 ±0.0161	0.8318 ±0.0114
Rel-MOSS	0.3571 ±0.0071	0.7999 ±0.0058	0.2372 ±0.0018	0.7153 ±0.0042	0.7419 ±0.0024	0.7959 ±0.0063	0.5308 ±0.0027	0.8524 ±0.0055

Table 2: Evaluation results comparison in terms of PR-AUC

\uparrow

and ROC-AUC

\uparrow

Dataset	f1-driver-top3	f1-driver-dnf	event-user-repeat	event-user-ignore
Standard RDL	1.0401	6.1067	2.9947	13.0044
Rel-MOSS	0.9979	5.7454	3.6566	14.0246
# Sample	1353	11411	3842	19239
Imb-ratio	4.86	7.36	1.04	4.92

Table 3: Practical runtime (s) comparison of RDL and Rel-MOSS.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
$\mathcal{R}$	Relational database (RDB)
$\mathcal{T}$	Set of tables in the RDB
$T,T_{t}$	A specific table in the relational database
$\mathcal{S}$	Set of linkages induced from the RDB foreign-primary key reference
$e,e_{\rm syn}$	RDB entity, synthetic entity
$e_{m}$	The $m$ -th feature of entity $e$

Appendix

Table 4: Notations and corresponding descriptions.

Dataset	f1-driver-top3	f1-driver-dnf	avito-user-clicks	avito-user-visits
# Positive	231	10046	2302	78467
# Negative	1122	1365	57152	8152
Imb-ratio	4.86	7.36	24.83	9.62
Dataset	event-user-repeat	event-user-ignore	stack-user-engagement	stack-user-badge
# Positive	1882	3247	68020	163048
# Negative	1960	15992	1292830	3223228

Appendix

Table 5: Extent of class imbalance for each dataset. Imb-ratio represents the imbalance ratio, which is defined as #

\mathtt{Majority}

\mathtt{Minority}

. The statistics is summarized on the train set, which determines the model training dynamics.

Figure 6: Performance of Rel-MOSS versus $\gamma$ in Eq. 13 on 6 datasets, in terms of balanced accuracy (B-Acc) and G-Mean.

Figure 7: Performance of Rel-MOSS versus $\omega$ in Eq. 11 on 6 datasets, in terms of balanced accuracy (B-Acc) and G-Mean.

Figure 8: Performance of Rel-MOSS versus learning rate $\eta$ on 6 datasets, in terms of balanced accuracy (B-Acc) and G-Mean.

Dataset	f1-driver-top3		f1-driver-dnf		event-user-repeat		event-user-ignore
Metric	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean
$d=32$	0.7345 ± 0.0127	0.7274 ± 0.0146	0.7184 ± 0.0195	0.7167 ± 0.0064	0.7317 ± 0.0246	0.7276 ± 0.0284	0.7263 ± 0.0072	0.7032 ± 0.0131
$d=64$	0.7886 ± 0.0134	0.7861 ± 0.0121	0.6717 ± 0.0071	0.6589 ± 0.0145	0.6991 ± 0.0273	0.6981 ± 0.0134	0.7101 ± 0.0020	0.6816 ± 0.0058
$d=128$	0.8098 ±0.0104	0.8014 ±0.0142	0.6510 ±0.0221	0.6262 ±0.0300	0.7411 ±0.0072	0.7398 ±0.0059	0.7460 ±0.0071	0.7362 ±0.0103
$d=256$	0.7122 ± 0.0119	0.7122 ± 0.0164	0.6755 ± 0.0281	0.6662 ± 0.0540	0.6864 ± 0.0249	0.6677 ± 0.0620	0.7210 ± 0.0158	0.6918 ± 0.0284
$d=512$	0.6203 ± 0.0195	0.5513 ± 0.0607	0.6830 ± 0.0301	0.6777 ± 0.0388	0.6778 ± 0.0216	0.6722 ± 0.0101	0.6988 ± 0.0150	0.6583 ± 0.0298

Appendix

Table 6: Entity classification performance of Rel-MOSS versus latent dimension

d

Dataset	f1-driver-top3		f1-driver-dnf		event-user-repeat		event-user-ignore
Metric	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean
$U=256$	0.7585 ± 0.0182	0.7582 ± 0.0128	0.6821 ± 0.0021	0.6811 ± 0.0141	0.7245 ± 0.0361	0.6894 ± 0.0257	0.7189 ± 0.0134	0.6889 ± 0.0179
$U=512$	0.7965 ± 0.0098	0.7839 ± 0.0378	0.6870 ± 0.0102	0.6789 ± 0.0126	0.6983 ± 0.0184	0.6379 ± 0.0446	0.7340 ± 0.0146	0.7244 ± 0.0274
$U=1024$	0.8098 ±0.0104	0.8014 ±0.0142	0.6510 ±0.0221	0.6262 ±0.0300	0.7411 ±0.0072	0.7398 ±0.0059	0.7460 ±0.0071	0.7362 ±0.0103
$U=2048$	0.7972 ± 0.0148	0.7959 ± 0.0135	0.6501 ± 0.0229	0.6332 ± 0.0351	0.7507 ± 0.0027	0.7205 ± 0.0169	0.7247 ± 0.0018	0.7038 ± 0.0056

Appendix

Table 7: Entity classification performance of Rel-MOSS versus memory bank scale

U

Dataset	f1-driver-top3		f1-driver-dnf		event-user-repeat		event-user-ignore
Metric	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean
$B=128$	0.6203 ± 0.0192	0.5513 ± 0.0120	0.7002 ± 0.0376	0.6969 ± 0.0538	0.6843 ± 0.0255	0.6824 ± ± 0.0248	0.7352 ± 0.0249	0.7314 ± 0.0364
$B=256$	0.8032 ± 0.0156	0.7984 ± 0.0261	0.6050 ± 0.0112	0.5465 ± 0.0424	0.7261 ± 0.0282	0.7190 ± 0.0442	0.7194 ± 0.0072	0.6886 ± 0.0137
$B=512$	0.8098 ±0.0104	0.8014 ±0.0142	0.6510 ±0.0221	0.6262 ±0.0300	0.7411 ±0.0072	0.7398 ±0.0059	0.7460 ±0.0071	0.7362 ±0.0103
$B=1024$	0.8185 ± 0.0124	0.8138 ± 0.0178	0.6879 ± 0.0132	0.6820 ± 0.0189	0.6576 ± 0.0084	0.6125 ± 0.0106	0.7085 ± 0.0073	0.6714 ± 0.0137

Appendix

Table 8: Entity classification performance of Rel-MOSS versus batch size

B

Dataset	f1-driver-top3		f1-driver-dnf		event-user-repeat		event-user-ignore
Metric	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean	B-Acc	G-Mean
RDL-HGT	0.5124 ±0.0175	0.1179 ±0.1670	0.5497 ±0.0324	0.3647 ±0.1116	0.6021 ±0.0737	0.4296 ±0.3045	0.6820 ±0.0059	0.6235 ±0.0147
Rel-MOSS (HGT)	0.7675 ±0.0870	0.7502 ±0.0832	0.6901 ±0.0458	0.6857 ±0.0624	0.6279 ±0.0640	0.5456 ±0.1788	0.7125 ±0.0078	0.6813 ±0.0119
RelGNN	0.4976 ±0.0034	0.0927 ±0.1311	0.5272 ±0.0066	0.2973 ±0.0332	0.6279 ±0.0671	0.5822 ±0.1229	0.6303 ±0.0418	0.5233 ±0.0869
Rel-MOSS (RelGNN)	0.7439 ±0.0514	0.7378 ±0.0702	0.7009 ±0.0486	0.7008 ±0.0342	0.6912 ±0.0266	0.6903 ±0.0262	0.7341 ±0.0208	0.7158 ±0.0320

Appendix

Table 9: Entity classification performance of Rel-MOSS with HGT and RelGNN backbones.

Explore similar work

What Makes a Desired Graph for Relational Deep Learning?
Jun 7, 2026Yao Cheng, Siqiang LuoGraph Structure Learning Graph Neural Networks
Nanyang Technological University, College of Computing and Data Science, Singapore.
A Fair Benchmarking of Deep Relational Database Learning Models
Jul 4, 2026Kazi F. Akhter, Bharath Ajendla, Manar D. SamadTransformer Graph Representation Learning
Department of Computer Science Tennessee State University Nashville, TN, USA · SAP AI Research and Innovation SAP Labs Palo Alto, CA, USA
Revisiting Simple Baselines for Relational Deep Learning
Oct 14, 2026Viet The Nguyen, Vijay Prakash Dwivedi, Gunther GustRelational Deep Learning
University of Würzburg · Seamless Labs, Inc. · Stanford University

Jun 7, 2026J/K move · Enter open · S save

What Makes a Desired Graph for Relational Deep Learning?

Yao Cheng, Siqiang Luo

Nanyang Technological University, College of Computing and Data Science, Singapore.

Graph Structure Learning Graph Neural Networks Relational Reasoning Relational Deep Learning Relational Databases

Open paper AskPDF arXiv

Relational deep learning (RDL) converts relational databases (RDBs) into heterogeneous graphs, but graphs derived directly from database schemas are often not well suited for how graph neural networks (GNNs) perform relational reasoning. We study what makes a relational graph suitable for deep learning and show that schema-derived graphs suffer from two systematic failures: information overload and semantic fragmentation. Our empirical analysis reveals that the desired graph is not the raw schema, but a result of controlled structural adaptation. Performance depends on balancing two operations: mitigating information overload via filtering, and repairing semantic fragmentation via injection. Specifically, filtering serves as a bias-variance knob with non-monotonic effects, while injection improves performance only when it explicitly restores the relational dependencies missing from the original schema. Based on these findings, we develop an end-to-end structural optimizer that applies both operations to adapt relational graphs automatically. Across 26 tasks spanning classification, regression, and recommendation, the optimized graphs consistently improve accuracy while often reducing inference cost.

Abstract

Explore similar work

What Makes a Desired Graph for Relational Deep Learning?

A Fair Benchmarking of Deep Relational Database Learning Models

Revisiting Simple Baselines for Relational Deep Learning