Hybrid BART-GNN Model Advances Relational Deep Learning
Researchers have developed an innovative hybrid model that integrates an enhanced BART encoder with a GraphSAGE-based graph neural network (GNN) for analyzing relational entity graphs (REGs). This advancement produced a ROC-AUC score of 67.40 in the driver-dnf task using the rel-f1 dataset. The model addresses shortcomings in traditional deep learning techniques that often oversimplify tables and overlook significant relational information. By utilizing REGs to depict databases, the GNN augments BART's row embeddings with pertinent relational context, surpassing previous performance benchmarks on RelBench. This work paves the way for creating foundational models tailored for relational data analysis.
Key facts
- Hybrid architecture combines fine-tuned BART encoder with GraphSAGE-based GNN over REGs
- Achieves ROC-AUC of 67.40 on driver-dnf task from rel-f1 dataset
- Conventional approaches flatten databases into single tables via manual feature engineering
- Relational deep learning (RDL) models databases as relational entity graphs (REGs)
- GNN substantially enriches BART's row embeddings with relational context
- Performance is competitive with supervised baselines on RelBench
- Proposes a path toward foundation models for relational databases
- Paper published on arXiv with ID 2605.16085
Entities
Institutions
- arXiv