deem-data/erase-bench

ERASE – A Real-World Aligned Benchmark for Unlearning in Recommender Systems

ERASE is a benchmark for evaluating machine unlearning methods in recommender systems under real-world inspired conditions. It provides a standardized framework for measuring unlearning efficiency, utility, and effectiveness across multiple datasets and models.

For full details, code, and documentation, see the GitHub repository.

Repository Contents

• saved/ — Trained, retrained, and unlearned model checkpoints
• logs/ — Logs from training, retraining, and unlearning runs, including efficiency, utility, and effectiveness metrics
• dataset/ — Instructions and scripts for downloading the datasets used in the benchmark