Research map / Graph learning / Knowledge and temporal graphs
Position: Relational Deep Learning - Graph Representation Learning on 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.
Position paper proposing Relational Deep Learning, which represents relational databases as temporal heterogeneous graphs and trains GNNs end-to-end, plus the RelBench benchmark with two databases and four tasks.
Problem
Machine learning on multi-table relational databases currently requires manual feature engineering (joins and aggregations) into a single table, which is slow, error-prone, leaks temporal information, and suboptimal.
Approach
View a relational database as a temporal heterogeneous graph with one node per row and edges from primary-foreign keys, then learn predictions end-to-end with message passing GNNs that only pass messages from earlier timestamps, with labels generated from historical data via training tables.
Key results
- No quantitative results are reported in this position paper; RelBench beta provides two databases (rel-amazon with 1.85M customers, 21.9M reviews; rel-stackex with 333K users, 415K posts) and four tasks.
Limitations
- No experimental results or baseline comparisons reported in the position paper
- Pre-extracted frozen entity-level encoders may be suboptimal for unusual data
- Stale links and distribution drift are unresolved
Experimental setup
Datasets: rel-amazon rel-stackex
Benchmarks: RelBench rel-amazon-ltv RelBench rel-amazon-churn RelBench rel-stackex-engage RelBench rel-stackex-votes
Metrics: MAE average precision
Baselines: XGBoost LightGBM
Builds on: message passing neural networks heterogeneous graph neural networks PyTorch Frame PyTorch Geometric XGBoost
Closest papers by meaning
- RelGNN: Composite Message Passing for Relational Deep Learning (ICML 2025 · similarity 0.89)
- RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases (NeurIPS 2025 · similarity 0.88)
- Griffin: Towards a Graph-Centric Relational Database Foundation Model (ICML 2025 · similarity 0.87)
- WikiDBs: A Large-Scale Corpus Of Relational Databases From Wikidata (NeurIPS 2024 · similarity 0.86)
- Joint Relational Database Generation via Graph-Conditional Diffusion Models (NeurIPS 2025 · similarity 0.86)
- 4DBInfer: A 4D Benchmarking Toolbox for Graph-Centric Predictive Modeling on RDBs (NeurIPS 2024 · similarity 0.85)
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.