Research map / Graph learning / Knowledge and temporal graphs
RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases
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 RDB2G-Bench, a precomputed benchmark of about 50k graph models of 5 RelBench databases across 12 tasks, used to evaluate 10 automatic RDB-to-graph modeling methods roughly 380x faster than on-the-fly GNN training.
Problem
Choosing how to convert a relational database into a graph (which tables/FKs to include, and whether rows become nodes or edges) strongly affects GNN performance, but evaluating candidate graph models requires costly repeated GNN training, so research on automatic RDB-to-graph modeling is hard to ca
Approach
Precompute downstream GNN performance, runtime and parameter counts for ~50k graph models over RelBench databases so that RDB-to-graph modeling methods can be benchmarked by table lookup instead of GNN training; benchmark heuristic, search-based and LLM-based methods on it.
Key results
- The best graph model yields up to about 10% higher performance than the common AR2N heuristic (abstract), with 11.3% spread on post-post-related and 4.83% gain on driver-top3.
- Precomputed metrics reduce benchmark evaluation time from over 850 GPU hours to 2.20 hours, about 380-389x faster.
- Most search baselines outperform AR2N with minimal exploration; Greedy Backward is best at small budgets and Bayesian Optimization overtakes it as budget grows.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Search graph models rather than use fixed AR2N modeling | RDB-to-graph modeling on rel-f1 driver-dnf · rel-f1 driver-dnf, 5% budget | AUC-ROC (%): 73.140 -> 74.557 | Helped |
| Search graph models rather than use fixed AR2N modeling | RDB-to-graph modeling on rel-f1 driver-top3 · rel-f1 driver-top3, 5% budget | AUC-ROC (%): 78.106 -> 81.879 | Helped |
| Search graph models rather than use fixed AR2N modeling | RDB-to-graph modeling on rel-f1 driver-position · rel-f1 driver-position, 5% budget | MAE: 3.9125 -> 3.8388 | Helped |
| Use precomputed graph-model performance metrics instead of on-the-fly GNN evaluation | RDB2G-Bench evaluation · RDB-to-graph method evaluation | speedup: 389× | Helped |
| Use Greedy Backward instead of more complex search under small budgets | RDB-to-graph modeling benchmark · 12 predictive tasks | average rank | Helped |
| Increase search budget for Bayesian Optimization | RDB-to-graph modeling benchmark · 12 predictive tasks | average rank | Helped |
| Use Chain-of-Thought prompting rather than the LLM baseline prompt | LLM baseline · RDB-to-graph modeling benchmark | performance | Helped |
| Use Row2Edge modeling | rel-event user-ignore · rel-event user-ignore | AUC-ROC | Hurt |
| Model event_attendees and user_friends rows as edges | rel-event user-repeat · rel-event user-repeat | AUC-ROC | Helped |
| Select fewer, relevant tables rather than include all tables and FK relationships | RDB graph modeling · multiple predictive tasks | predictive performance | Helped |
Surprising findings
- Greedy and random methods perform comparably to advanced Bayesian Optimization and reinforcement learning approaches.
- Cross-task Spearman correlations are mostly below 0.4, so a graph model effective for one task rarely transfers to others on the same RDB.
Limitations
- Graph model space is limited by constraints: tables with more than two FKs cannot be modeled as edges (hyperedges out of scope).
- Only five RelBench databases and twelve tasks are covered.
- GPS neighbor sampling reduced from 128 to 32 to fit a single GPU, and edge-modeled graphs use only 12 sampled neighbors.
Experimental setup
Datasets: RelBench rel-avito RelBench rel-event RelBench rel-f1 RelBench rel-stack RelBench rel-trial
Benchmarks: RDB2G-Bench
Metrics: AUC-ROC MAE MAP MSE rank averaged over tasks Spearman rank correlation evaluation time
Baselines: Random AR2N Greedy Forward Greedy Backward Greedy Local Evolutionary Algorithm Bayesian Optimization Reinforcement Learning LLM baseline LLM-CoT baseline
Builds on: RelBench PyTorch Frame BANANAS AutoG GraphSAGE
Closest papers by meaning
- 4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs (NeurIPS 2024 · similarity 0.92)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.90)
- Joint Relational Database Generation via Graph-Conditional Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024 · similarity 0.88)
- TGB 2.0: A Benchmark for Learning on Temporal Knowledge Graphs and Heterogeneous Graphs (NeurIPS 2024 · similarity 0.87)
- AutoG: Towards automatic graph construction from tabular data (ICLR 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.