This is the official implementation of ReFuGe (Feature Generation for Prediction Tasks on Relational Databases with LLM Agents).

Accepted in ACM WWW 2026

ReFuGe requires core dependencies:

RelBench library (required) — used for loading RDB benchmark datasets, schemas, and task definitions.

Install RelBench via:

pip install relbench

For details, please refer to https://github.com/snap-stanford/relbench

Claude API — ReFuGe uses Claude models as the backbone LLM agents (schema selection, feature generation, and feature filtering). Make sure to set your API key:

export ANTHROPIC_API_KEY="YOUR_API_KEY"

Each dataset–task pair comes with its own Python script. To run experiments, simply execute the corresponding file (e.g., amazonchurn.py etc.).

python {dataset}{task}.py