Revisiting Simple Baselines for Relational Deep Learning

Viet The Nguyen; Vijay Prakash Dwivedi; Gunther Gust

cs.LGOct 14, 2026

Revisiting Simple Baselines for Relational Deep Learning

Authors: Viet The Nguyen, Vijay Prakash Dwivedi, Gunther Gust

Organizations: University of Würzburg · Seamless Labs, Inc. · Stanford University

Abstract

Relational deep learning has emerged as a powerful paradigm for learning end-to-end predictive models over relational databases by representing them as temporal heterogeneous graphs. Recent advances based on graph neural networks and Transformers have achieved strong performance across diverse domains. However, the effectiveness of simpler relational baselines remains underexplored, making it difficult to assess how much architectural complexity is needed to achieve strong predictive performance. Motivated by the observation that predictive signals arise from target attributes, relational attributes, and historical labels, we revisit how effectively these signals can be exploited using simple propagation methods. We adapt fixed low-pass feature propagation to relational entity graphs and train an MLP for prediction. The resulting predictions are further refined using Correct-and-Smooth, a graph-based label propagation algorithm that incorporates historical labels over task-specific meta-graphs connecting target entities. Across classification and regression tasks on RelBench, an MLP trained on neighborhood-filtered features outperforms one using only target-entity features on most tasks and recovers a substantial portion of the performance of heterogeneous GNNs. Correct-and-Smooth provides further gains from historical labels, although applying it requires a separately constructed meta-graph because connections among target entities are not directly specified by the relational entity graph. Together, our studies revisit the effectiveness of established propagation methods as simple RDL baselines while identifying canonical meta-graph construction as a key challenge for effective label propagation.

Figures & tables

Figure 1: Example of graph conversion. (Left) A relational entity graph, where each node is an entity, and edges are primary–foreign key links. (Right) All nodes are collapsed into a single type, and feature sets are merged and normalized.

Task	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)	HeteroGNN
RelBench-v1
amazon / user-churn	$53.65{\scriptstyle\pm 0.18}$	$68.12{\scriptstyle\pm 0.05}$	$\mathbf{68.69}{\scriptstyle\pm 0.04}$	$\mathbf{70.42}{\scriptstyle\pm 0.05}$
amazon / item-churn	$66.60{\scriptstyle\pm 0.44}$	$80.58{\scriptstyle\pm 0.08}$	$\mathbf{81.03}{\scriptstyle\pm 0.07}$	$\mathbf{82.81}{\scriptstyle\pm 0.03}$
avito / user-clicks	$51.74{\scriptstyle\pm 0.51}$	$65.08{\scriptstyle\pm 1.06}$	$\mathbf{65.10}{\scriptstyle\pm 1.06}$	$\mathbf{65.90}{\scriptstyle\pm 1.95}$
avito / user-visits	$49.64{\scriptstyle\pm 0.01}$	$\mathbf{66.29}{\scriptstyle\pm 0.20}$	$\mathbf{66.41}{\scriptstyle\pm 0.18}$	$66.20{\scriptstyle\pm 0.10}$
hm / user-churn	$55.25{\scriptstyle\pm 0.05}$	$68.21{\scriptstyle\pm 0.06}$	$\mathbf{69.16}{\scriptstyle\pm 0.05}$	$\mathbf{69.88}{\scriptstyle\pm 0.21}$

Table 1: Entity classification on RelBench-v1 and RelBench-v2 (AUROC

\uparrow

Task	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)	HeteroGNN
RelBench-v1
amazon / user-ltv	$16.757{\scriptstyle\pm 0.007}$	$\mathbf{14.656}{\scriptstyle\pm 0.013}$	$\mathbf{14.656}{\scriptstyle\pm 0.013}$	$\mathbf{14.313}{\scriptstyle\pm 0.013}$
amazon / item-ltv	$59.958{\scriptstyle\pm 0.150}$	$\mathbf{50.994}{\scriptstyle\pm 0.168}$	$\mathbf{50.994}{\scriptstyle\pm 0.168}$	$\mathbf{50.053}{\scriptstyle\pm 0.163}$
avito / ad-ctr	$0.043{\scriptstyle\pm 0.000}$	$\mathbf{0.040}{\scriptstyle\pm 0.003}$	$\mathbf{0.040}{\scriptstyle\pm 0.003}$	$\mathbf{0.041}{\scriptstyle\pm 0.001}$
hm / item-sales	$0.076{\scriptstyle\pm 0.000}$	$\mathbf{0.060}{\scriptstyle\pm 0.000}$	$\mathbf{0.060}{\scriptstyle\pm 0.000}$	$\mathbf{0.056}{\scriptstyle\pm 0.000}$
stack / post-votes	$0.068{\scriptstyle\pm 0.000}$	$\mathbf{0.066}{\scriptstyle\pm 0.000}$	$\mathbf{0.066}{\scriptstyle\pm 0.000}$	$\mathbf{0.065}{\scriptstyle\pm 0.000}$

Table 2: Entity regression on RelBench-v1 and RelBench-v2 (MAE

\downarrow

Figure 2: Trainable parameters (left) and MLP parameter reduction relative to HeteroGNN (right).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Train	Val	Test	Seed times
RelBench-v1
classification
amazon / user-churn	4,708,383	409,792	351,885	33
amazon / item-churn	2,536,014	177,689	166,842	33
avito / user-clicks	59,454	21,183	47,996	5
avito / user-visits	86,619	29,979	36,129	5

Appendix

Table 3: Task sizes by official split and number of distinct seed times.

Database	Tables	Nodes	Input dim	Text cols	Num. cols	Other	MLP params
RelBench-v1
rel-amazon	3	15,000,713	1,814	6	7	2	464,897
rel-avito	8	20,679,117	622	2	40	0	159,745
rel-hm	3	16,664,809	3,092	10	13	14	792,065
rel-stack	7	4,247,264	4,529	15	29	6	1,159,937
rel-trial	15	5,434,924	15,145	50	53	30	3,877,633

Appendix

Table 4: Graph and feature statistics by database. Input dimensions include degree and age, used only for

K>0

Task	Filtering	Training MLP	HeteroGNN
RelBench-v1
classification
amazon / user-churn	$8.5$	$16.2$	$28.8$
amazon / item-churn	$9.2$	$10.2$	$15.3$
avito / user-clicks	$1.1$	$0.3$	$9.6$
avito / user-visits	$1.8$	$0.3$	$3.9$

Appendix

Table 5: Filtering and training time per task (minutes); timing protocol in Appendix B.2 .

Panel of Figure 3 — Figure 3: Meta-graphs for $(e,t)$ with undirected edges and labels observed before $t$ . Own histories lie left of $e$ , related histories right; indices increase from newest to oldest, with $y_{j,k}$ denoting related entity $v_{j}$ ’s $k$ th label. Dashed relays are unlabeled intermediate nodes that carry base predictions and connect historical labels to the seed in two hops. In (e), $r_{1},\ldots,r_{m}$ represent all $m$ own past labels, each with separate copies of up to three newest own labels. Table 6 gives the history limits.

Variant	Panel	Own history	Related history (per entity)
Self	(a)	Star, all labels	None
Cross	(b)	None	Most recent label, directly
Cross+Self	(c)	Star, all labels	Most recent label, directly
Cross+Hist	(d)	None	Relay, up to 3 labels
All	(e)	One relay per past label; up to 3 label copies each	Relay, up to 3 labels
Self-Chain	(f)	Chain, all labels	None

Appendix

Table 6: Meta-graph variants in Figure 3 . A star attaches labels directly; a chain orders them newest first; a relay is an unlabeled intermediate node between the seed and historical labels.

	Self	Cross	Cross+Self	Cross+Hist	All	Self-Chain	Chain+Cross	Chain+Hist-3	Chain+Hist
RelBench-v1
classification (AUROC $\times$ 100 $\uparrow$ )
amazon / user-churn	$68.42{\scriptstyle\pm 0.05}$	$68.47{\scriptstyle\pm 0.04}$	$68.69{\scriptstyle\pm 0.04}$	$68.44{\scriptstyle\pm 0.04}$	$68.61{\scriptstyle\pm 0.04}$	$68.47{\scriptstyle\pm 0.05}$	$68.59{\scriptstyle\pm 0.04}$	$\mathbf{68.77}{\scriptstyle\pm 0.05}$	$\mathbf{68.76}{\scriptstyle\pm 0.05}$
amazon / item-churn	$80.93{\scriptstyle\pm 0.06}$	$81.02{\scriptstyle\pm 0.06}$	$81.03{\scriptstyle\pm 0.07}$	$81.07{\scriptstyle\pm 0.06}$	$\mathbf{81.09}{\scriptstyle\pm 0.06}$	$80.76{\scriptstyle\pm 0.06}$	$81.07{\scriptstyle\pm 0.06}$	$\mathbf{81.18}{\scriptstyle\pm 0.05}$	$\mathbf{81.18}{\scriptstyle\pm 0.05}$
avito / user-clicks	$62.34{\scriptstyle\pm 0.85}$	$65.08{\scriptstyle\pm 1.06}$	$\mathbf{65.10}{\scriptstyle\pm 1.06}$	$\mathbf{65.44}{\scriptstyle\pm 1.04}$	$64.86{\scriptstyle\pm 1.02}$	$61.91{\scriptstyle\pm 0.77}$	$64.62{\scriptstyle\pm 1.09}$	$65.09{\scriptstyle\pm 1.06}$	$65.09{\scriptstyle\pm 1.06}$
avito / user-visits	$66.43{\scriptstyle\pm 0.17}$	$66.40{\scriptstyle\pm 0.15}$	$66.41{\scriptstyle\pm 0.18}$	$66.36{\scriptstyle\pm 0.18}$	$\mathbf{66.72}{\scriptstyle\pm 0.11}$	$66.40{\scriptstyle\pm 0.17}$	$\mathbf{66.45}{\scriptstyle\pm 0.14}$	$66.37{\scriptstyle\pm 0.18}$	$66.37{\scriptstyle\pm 0.18}$

Appendix

Table 7: Label-graph variants with Optuna tuning on the trained filtered MLP over 10 seeds.

Task	Best variant	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)
amazon / user-churn	Chain+Hist-3	$53.65{\scriptstyle\pm 0.18}$	$\mathbf{68.12}{\scriptstyle\pm 0.05}$	$\mathbf{68.77}{\scriptstyle\pm 0.05}$
amazon / item-churn	Chain+Hist-3	$66.60{\scriptstyle\pm 0.44}$	$\mathbf{80.58}{\scriptstyle\pm 0.08}$	$\mathbf{81.18}{\scriptstyle\pm 0.05}$
avito / user-clicks	Cross+Hist	$51.74{\scriptstyle\pm 0.51}$	$\mathbf{65.08}{\scriptstyle\pm 1.06}$	$\mathbf{65.44}{\scriptstyle\pm 1.04}$
avito / user-visits	All	$49.64{\scriptstyle\pm 0.01}$	$\mathbf{66.29}{\scriptstyle\pm 0.20}$	$\mathbf{66.72}{\scriptstyle\pm 0.11}$
hm / user-churn	Cross+Self	$55.25{\scriptstyle\pm 0.05}$	$\mathbf{68.21}{\scriptstyle\pm 0.06}$	$\mathbf{69.16}{\scriptstyle\pm 0.05}$
stack / user-badge	Cross+Self	$51.93{\scriptstyle\pm 1.87}$	$\mathbf{86.04}{\scriptstyle\pm 0.08}$	$\mathbf{86.16}{\scriptstyle\pm 0.09}$

Appendix

Table 8: Best meta-graph per task (test AUROC

\times 100

); reporting as in Table 1 .

Task	Selected construction	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)
amazon / user-ltv	Self-Chain	$16.757{\scriptstyle\pm 0.007}$	$\mathbf{14.656}{\scriptstyle\pm 0.013}$	$\mathbf{14.587}{\scriptstyle\pm 0.009}$
amazon / item-ltv	Self-Chain	$59.958{\scriptstyle\pm 0.150}$	$\mathbf{50.994}{\scriptstyle\pm 0.168}$	$\mathbf{46.932}{\scriptstyle\pm 0.111}$
avito / ad-ctr	Self-Chain	$0.043{\scriptstyle\pm 0.000}$	$\mathbf{0.040}{\scriptstyle\pm 0.003}$	$\mathbf{0.038}{\scriptstyle\pm 0.001}$
hm / item-sales	Self-Chain	$0.076{\scriptstyle\pm 0.000}$	$\mathbf{0.060}{\scriptstyle\pm 0.000}$	$\mathbf{0.044}{\scriptstyle\pm 0.000}$
stack / post-votes	Self	$\mathbf{0.068}{\scriptstyle\pm 0.000}$	$\mathbf{0.066}{\scriptstyle\pm 0.000}$	$\mathbf{0.066}{\scriptstyle\pm 0.000}$
trial / site-success	Chain+Hist-3	$\mathbf{0.426}{\scriptstyle\pm 0.002}$	$0.437{\scriptstyle\pm 0.006}$	$\mathbf{0.425}{\scriptstyle\pm 0.006}$

Appendix

Table 9: Regression with validation-selected meta-graphs.

Task	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)	HeteroGNN	RelGNN	RDBLearn
RelBench-v1
classification (AUROC $\times$ 100 $\uparrow$ )
amazon / user-churn	$53.65{\scriptstyle\pm 0.18}$	$68.12{\scriptstyle\pm 0.05}$	$68.69{\scriptstyle\pm 0.04}$	$\mathbf{70.42}{\scriptstyle\pm 0.05}$	$\mathbf{70.99}$	$67.57$
amazon / item-churn	$66.60{\scriptstyle\pm 0.44}$	$80.58{\scriptstyle\pm 0.08}$	$81.03{\scriptstyle\pm 0.07}$	$\mathbf{82.81}{\scriptstyle\pm 0.03}$	$\mathbf{82.64}$	$82.07$
avito / user-clicks	$51.74{\scriptstyle\pm 0.51}$	$65.08{\scriptstyle\pm 1.06}$	$65.10{\scriptstyle\pm 1.06}$	$65.90{\scriptstyle\pm 1.95}$	$\mathbf{68.23}$	$\mathbf{69.04}$
avito / user-visits	$49.64{\scriptstyle\pm 0.01}$	$\mathbf{66.29}{\scriptstyle\pm 0.20}$	$\mathbf{66.41}{\scriptstyle\pm 0.18}$	$66.20{\scriptstyle\pm 0.10}$	$66.18$	$65.49$

Appendix

Table 10: Extended results with more advanced models.

		Without structural columns		With structural columns
Task	MLP (target attributes)	MLP (+relational attributes)	C&S (+historical labels)	MLP (+relational attributes)	C&S (+historical labels)
RelBench-v1
classification (AUROC $\times$ 100 $\uparrow$ )
amazon / user-churn	$53.65{\scriptstyle\pm 0.18}$	$65.82{\scriptstyle\pm 0.10}$	$67.35{\scriptstyle\pm 0.04}$	$\mathbf{68.12}{\scriptstyle\pm 0.05}$	$\mathbf{68.69}{\scriptstyle\pm 0.04}$
amazon / item-churn	$66.60{\scriptstyle\pm 0.44}$	$75.91{\scriptstyle\pm 0.07}$	$78.07{\scriptstyle\pm 0.03}$	$\mathbf{80.58}{\scriptstyle\pm 0.08}$	$\mathbf{81.03}{\scriptstyle\pm 0.07}$
avito / user-clicks	$51.74{\scriptstyle\pm 0.51}$	$\mathbf{66.26}{\scriptstyle\pm 0.50}$	$\mathbf{66.26}{\scriptstyle\pm 0.50}$	$65.08{\scriptstyle\pm 1.06}$	$\mathbf{65.10}{\scriptstyle\pm 1.06}$

Appendix

Table 11: Ablation study on structural columns (degree and age).

Panel of Figure 4 — Figure 4: Effect of the filter depth per task.

Task	MLP (+relational attributes)	+ self hist.	+ (self+cross) hist.	+ C&S
amazon / user-churn	$68.09$	$68.51$	$\mathbf{69.05}$	$\mathbf{69.30}$
amazon / item-churn	$80.58$	$80.91$	$\mathbf{81.12}$	$\mathbf{81.41}$
avito / user-clicks	$\mathbf{65.35}$	$65.05$	$\mathbf{65.46}$	$65.03$
avito / user-visits	$66.11$	$\mathbf{66.55}$	$\mathbf{66.68}$	$\mathbf{66.68}$
hm / user-churn	$67.84$	$\mathbf{69.65}$	$\mathbf{69.76}$	$\mathbf{69.76}$
stack / user-badge	$86.06$	$86.26$	$\mathbf{86.29}$	$\mathbf{86.34}$

Appendix

Table 12: Historical labels as features and with C&S.

Explore similar work

Jun 7, 2026

What Makes a Desired Graph for Relational Deep Learning?

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.

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

Ask PDF

May 15, 2026

Towards Foundation Models for Relational Databases with Language Models and Graph Neural Networks

Relational databases store much of the world's structured information, and they are essential for driving complex predictive applications. However, deep learning progress on relational data remains limited, as conventional approaches flatten databases into single tables via manual feature engineering, discarding relational context. Relational deep learning (RDL) addresses this by modeling databases as relational entity graphs (REGs) for graph neural networks (GNNs), but remains task- and database-specific. To combine the strengths of both paradigms, we propose a hybrid architecture combining a fine-tuned BART encoder to capture intra-row semantics with a GraphSAGE-based GNN over REGs to inject relational context. Experiments on RelBench show that the GNN substantially enriches BART's row embeddings, achieving a ROC-AUC of 67.40 on the driver-dnf task from the rel-f1 dataset. This performance is competitive with supervised baselines such as LightGBM (68.86) and narrows the gap to RDL (72.62) to within 5.22 points, though a substantial gap remains to state-of-the-art foundation models such as KumoRFM (82.63). These results suggest that lightweight hybrid LM-GNN architectures offer a promising and resource-efficient path towards foundation models for relational databases.

Jingcheng Wu, Ratan Bahadur Thapa, Mojtaba Nayyeri +4

University of Stuttgart, Stuttgart, Germany · Web and Internet Science Research Group, University of Southampton, Southampton, United Kingdom

Relational Foundation Models Tabular Foundation Models Graph Neural Networks Graph Foundation Models Relational Deep Learning Relational Databases

Ask PDF

Jul 4, 2026

A Fair Benchmarking of Deep Relational Database Learning Models

Relational databases (RDBs) are the primary data infrastructure in many enterprises, yet recent deep learning methods designed for RDBs have been evaluated under inconsistent experimental protocols, making fair comparison difficult. We present one of the first systematic benchmarking studies of recently released deep learning methods for RDBs, evaluating them across five relational databases, with one classification and one regression task for each. We refactor all deep RDB models to allow the full range of experimental procedures to be applied consistently across all methods. Our findings indicate that the relational transformer (RT) approach delivers the strongest overall performance on both classification and regression tasks compared to the state-of-the-art graph-based modeling and learning of RDBs. Even for single-table learning tasks, deep learning methods designed for RDBs outperform the leading tabular foundation model, TabPFN 2.5. Extending learning from a single table (hop = 0) to multiple tables (hop = 1, 2) by connecting neighboring tables in relational databases enhances performance, but the additional benefit from higher hops diminishes as computational overhead grows. Deep RDB learning methods have the potential to challenge state-of-the-art tabular foundation models, especially on large-scale enterprise data. The source code for this benchmarking study is publicly available.

Kazi F. Akhter, Bharath Ajendla, Manar D. Samad

Department of Computer Science Tennessee State University Nashville, TN, USA · SAP AI Research and Innovation SAP Labs Palo Alto, CA, USA

Transformer Graph Representation Learning Relational Deep Learning Tabular Prediction Relational Databases

Ask PDF