Research map / Graph learning / Knowledge and temporal graphs
4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces 4DBInfer, an open-source toolbox and benchmark suite (8 large-scale RDB datasets, 15 tasks) for graph-centric predictive modeling on relational databases, comparing single-table, join, DFS late-fusion and GNN early-fusion baselines across graph extraction methods.
Problem
Predictive ML on relational databases lacks public, large-scale, native multi-table benchmarks; existing evaluations rely on pre-processed graphs or single-table data, so tabular and GNN approaches cannot be fairly compared.
Approach
Frames RDB prediction as a sampling operator over graphs extracted from RDBs (Row2Node, Row2N/E, with dummy tables) followed by either late fusion (DFS features plus tabular models) or early fusion (heterogeneous GNNs), and evaluates the full 4D grid of datasets, tasks, graph extractors and predictors on native RDB benchmarks.
Key results
- GNNs (early fusion) generally outperform DFS late fusion and join baselines; e.g., on AVS-Retention, R-GCN with Row2Node reaches 0.5578 vs DFS MLP 0.5690 and Join MLP 0.5618 AUC, while on AB-Churn R-PNA reaches 0.7645 vs DFS best 0.7291 AUC.
- Graph extraction choice matters: Row2N/E wins on OB-CTR, AB-Purchase, MAG-Venue and MAG-Cite while Row2Node wins on DN-CTR, AB-Churn and SE-Popularity.
- Removing dummy tables drops AVS-Retention R-GCN AUC from 0.5653 to 0.4761.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GNNs instead of late-fusion DFS | 4DBInfer predictive models · Foreign-key prediction tasks | MRR | Helped |
| Use early-fusion GNNs instead of DFS | 4DBInfer predictive models · Across benchmarks | task-specific evaluation metric | Mixed |
| Use Row2N/E graph extraction instead of Row2Node | GNN baselines · OB CTR, AB Purchase, MAG Venue and MAG Citation | task-specific evaluation metric | Helped |
| Use Row2Node graph extraction instead of Row2N/E | GNN baselines · DN CTR, AB Churn and SE Popularity | task-specific evaluation metric | Helped |
| Add dummy tables (ablation) | DFS MLP · AB Rating | RMSE: 1.0291 -> 0.9847 | Helped |
| Add dummy tables (ablation) | DFS MLP · AVS Retention | AUC: 0.5469 -> 0.5690 | Helped |
| Add dummy tables (ablation) | R-GCN (Row2Node) · AVS Retention | AUC: 0.4761 -> 0.5653 | Helped |
| Add dummy tables (ablation) | GNN models · OB CTR | AUC | No clear effect |
| Add observable labels for label propagation (ablation) | 4DBInfer label-propagation models · OB CTR | AUC: over 0.10 | Helped |
| Use neural common neighbors (NCN) (ablation) | GNN baselines · 7 of 8 key or relationship attribute prediction benchmarks | task-specific evaluation metric | Helped |
Surprising findings
- Simple join-based tabular baselines can be competitive with or outperform GNNs on some smaller prior benchmarks, but not on the larger 4DBInfer datasets.
- DFS late fusion remains a strong competitor on AVS-Retention, DN-CTR and SE-Popularity.
Limitations
- No single baseline dominates across all scenarios, so conclusions are dataset- and task-dependent.
- Some benchmarks lack features in the target table, so single-table models are N/A on several tasks.
- Some datasets were down-sampled to around 100K samples, and several benchmarks use hand-chosen negative samples.
Experimental setup
Datasets: AVS Outbrain Diginetica RetailRocket Amazon Book Reviews StackExchange MAG Seznam
Benchmarks: AVS Outbrain Diginetica RetailRocket Amazon Book Reviews StackExchange MAG Seznam
Metrics: AUC MRR RMSE accuracy
Baselines: MLP DeepFM FT-Transformer XGBoost AutoGluon R-GCN R-GAT HGT R-PNA
Builds on: Deep Feature Synthesis R-GCN HGT FT-Transformer GraphBolt
Closest papers by meaning
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.92)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.88)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.87)
- GraphLand: Evaluating Graph Machine Learning Models on Diverse Industrial Data (NeurIPS 2025 · similarity 0.86)
- TabArena: A Living Benchmark for Machine Learning on Tabular Data (NeurIPS 2025 · similarity 0.86)
- Joint Relational Database Generation via Graph-Conditional Diffusion Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.