RDB2G-Bench

This is an offical dataset of the paper RDB2G-Bench: A Comprehensive Benchmark for Automatic Graph Modeling of Relational Databases.

RDB2G-Bench is a toolkit for benchmarking graph-based analysis and prediction tasks by converting relational database data into graphs.

Our code is available at GitHub.

Overview

RDB2G-Bench provides comprehensive performance evaluation data for graph neural network models applied to relational database tasks. The dataset contains extensive experiments across multiple graph configurations and architectures.

Dataset Summary

Each CSV file contains experimental results with the following columns:

Column	Description
idx	Unique identifier for each experimental configuration
graph	Binary-encoded string representing the graph structure configuration (e.g., "graph_00000000000010")
train_metric	Performance metric on the training set (e.g., AUC for classification, MSE for regression, MAP for recommendation)
valid_metric	Performance metric on the validation set
test_metric	Performance metric on the test set
params	Total number of trainable parameters in the model
train_time	Training time in seconds
valid_time	Validation time in seconds
test_time	Testing time in seconds
dataset	Name of the RDB dataset used (e.g., "rel-avito", "rel-f1")
task	Name of the prediction task (e.g., "ad-ctr", "user-clicks")
seed	Random seed for reproducibility

Graph Column Specification

The graph column uses binary encoding to represent different edge configurations in the graph structure. Each bit position corresponds to a specific edge type as defined in edge_info.json:

• 1: Edge is connected
• 0: Edge is disconnected

Edge Types:

• f2p: Standard foreign key relationships that each table row is transformed into a node (Row2Node).
• r2e: Converted relationships that each table row is transformed into an edge (Row2N/E).

The complete edge mapping for each dataset can be found in the edge_info.json file.

Reference

The dataset construction and implementation of RDB2G-Bench based on RelBench framework.

License

This project is distributed under the MIT License as specified in the LICENSE file.