Overview

Relational Deep Learning is a new approach for end-to-end representation learning on data spread across multiple tables, such as in a relational database (see our position paper). Relational databases are the world’s most widely used data management system, and are used for industrial and scientific purposes across many domains. RelBench is a benchmark designed to facilitate efficient, robust and reproducible research on end-to-end deep learning over relational databases.

RelBench contains 7 realistic, large-scale, and diverse relational databases spanning domains including medical, social networks, e-commerce and sport. Each database has multiple predictive tasks (30 in total) defined, each carefully scoped to be both challenging and of domain-specific importance. It provides full support for data downloading, task specification and standardized evaluation in an ML-framework-agnostic manner.

Additionally, RelBench provides a first open-source implementation of a Graph Neural Network based approach to relational deep learning. This implementation uses PyTorch Geometric to load the data as a graph and train GNN models, and PyTorch Frame for modeling tabular data. Finally, there is an open leaderboard for tracking progress.

Key Papers

RelBench: A Benchmark for Deep Learning on Relational Databases

This paper details our approach to designing the RelBench benchmark. It also includes a key user study showing that relational deep learning can produce performant models with a fraction of the manual human effort required by typical data science pipelines. This paper is useful for a detailed understanding of RelBench and our initial benchmarking results. If you just want to quickly familiarize with the data and tasks, the website is a better place to start.

Position: Relational Deep Learning - Graph Representation Learning on Relational Databases (ICML 2024)

This paper outlines our proposal for how to do end-to-end deep learning on relational databases by combining graph neural networsk with deep tabular models. We reccomend reading this paper if you want to think about new methods for end-to-end deep learning on relational databases. The paper includes a section on possible directions for future research to give a snapshot of some of the research possibilities there are in this area.

Design of RelBench

RelBench has the following main components:

1. 7 databases with a total of 30 tasks; both of these automatically downloadable for ease of use
2. Easy data loading, and graph construction from pkey-fkey links
3. Your own model, which can use any deep learning stack since RelBench is framework-agnostic. We provide a first model implementation using PyTorch Geometric and PyTorch Frame.
4. Standardized evaluators - all you need to do is produce a list of predictions for test samples, and RelBench computes metrics to ensure standardized evaluation
5. A leaderboard you can upload your results to, to track SOTA progress.

Installation

You can install RelBench using pip:

pip install relbench

This will allow usage of the core RelBench data and task loading functionality.

To additionally use relbench.modeling, which requires PyTorch, PyTorch Geometric and PyTorch Frame, install these dependencies manually or do:

pip install relbench[full]

For the scripts in the examples directory, use:

pip install relbench[example]

Then, to run a script:

git clone https://github.com/snap-stanford/relbench
cd relbench/examples
python gnn_node.py --dataset rel-f1 --task driver-position

Package Usage

This section provides a brief overview of using the RelBench package. For a more in-depth coverage see the Tutorials section. For detailed documentations, please see the code directly.

Imports:

from relbench.base import Table, Database, Dataset, EntityTask
from relbench.datasets import get_dataset
from relbench.tasks import get_task

Get a dataset, e.g., rel-amazon:

dataset: Dataset = get_dataset("rel-amazon", download=True)

dataset consists of a Database object and temporal splitting times dataset.val_timestamp and dataset.test_timestamp.

To get the database:

db: Database = dataset.get_db()

Various tasks can be defined on a dataset. For example, to get the user-churn task for rel-amazon:

task: EntityTask = get_task("rel-amazon", "user-churn", download=True)

A task provides train/val/test tables:

train_table: Table = task.get_table("train")
val_table: Table = task.get_table("val")
test_table: Table = task.get_table("test")

You can build your model on top of the database and the task tables. After training and validation, you can make prediction from your model on the test table. Suppose your prediction test_pred is a NumPy array following the order of task.test_table, you can call the following to get the evaluation metrics:

task.evaluate(test_pred)

Additionally, you can evaluate validation (or training) predictions as such:

task.evaluate(val_pred, val_table)

Tutorials

Notebook	Try on Colab	Description
load_data.ipynb		Load and explore RelBench data
train_model.ipynb		Train your first GNN-based model on RelBench
custom_dataset.ipynb		Use your own data in RelBench
custom_task.ipynb		Define your own ML tasks in RelBench

Contributing

Please check out CONTRIBUTING.md if you are interested in contributing datasets, tasks, bug fixes, etc. to RelBench.

Cite RelBench

If you use RelBench in your work, please cite our position and benchmark papers:

@inproceedings{rdl,
  title={Position: Relational Deep Learning - Graph Representation Learning on Relational Databases},
  author={Fey, Matthias and Hu, Weihua and Huang, Kexin and Lenssen, Jan Eric and Ranjan, Rishabh and Robinson, Joshua and Ying, Rex and You, Jiaxuan and Leskovec, Jure},
  booktitle={Forty-first International Conference on Machine Learning}
}

@misc{relbench,
      title={RelBench: A Benchmark for Deep Learning on Relational Databases},
      author={Joshua Robinson and Rishabh Ranjan and Weihua Hu and Kexin Huang and Jiaqi Han and Alejandro Dobles and Matthias Fey and Jan E. Lenssen and Yiwen Yuan and Zecheng Zhang and Xinwei He and Jure Leskovec},
      year={2024},
      eprint={2407.20060},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2407.20060},
}