---

If you find our work useful, please consider giving us a star🌟

📝 TODO

• Code Release
• Paper Release
• Datasets

📋 Table of Contents

• Customized Retrieval-Augmented Generation with LLM for Debiasing Recommendation Unlearning
  • 🔎 Overview
  • 🛠️ Installation
  • 🚀 Quick Start
  • ⚙️ Citation

Overview

CRAGRU is a unified framework that integrates RAG (Retrieval-Augmented Generation), Large Language Models (LLMs), and Recommendation Unlearning.
It enables:

• User-level and item-level unlearning
• Debiasing through controlled prompt design
• LLM-based recommendation generation
• Comparison and fusion with traditional recommender models
• Dataset clustering, DP strategy exploration, knapsack optimization, and more

The framework is modular, reproducible, and designed for flexible experimentation.

The framework of CRAGRU.

📦 Key Features

🔍 RAG-Enhanced LLM Recommendation Structured prompt design ensures controlled and interpretable LLM reasoning.

🧹 Efficient Recommendation Unlearning Supports flexible removal of user interactions or item histories.

📈 Dataset Analytics Suite Includes clustering, statistical analysis, and knapsack-based optimization.

🧩 Modular Architecture Every stage can be swapped or extended easily for research purposes.

Installation

DRAGRU works with the following operating systems:

• Linux
• Windows 10
• macOS X

DRAGRU requires Python version 3.10.12 or later.

DRAGRU requires torch version 2.5.1 or later. If you want to use DRAGRU with GPU,

Install

pip install -r requirements.txt

Download GoogleNews-vectors-negative300.bin and put it in the library file of your python directory

🚀 Quick Start

Below is the complete DRAGRU workflow, including one-sentence explanations and directly runnable commands.

---

1️⃣ Split Forget / Remain Sets

Description: Splits the dataset into the forget set and remain set, which serve as the foundation for all downstream unlearning tasks.

python DRAGRU/movie-lens/dataset_split.py

2️⃣ Item Clustering

Description: Performs item clustering using K-means + Word2Vec to provide semantic grouping for DP strategies and prompt construction.

python DRAGRU/movie-lens/statistics/item_cluster.py

3️⃣ Construct LLM Prompts

Description: Creates prompt files based on the remain set, serving as structured input for LLM-based recommendation.

python DRAGRU/movie-lens/data_preprocess_unlearning.py

4️⃣ Run LLM Recommendation

Description: Generates recommendation results using a large language model, with optional fallback to traditional models.

python DRAGRU/movie-lens/llm_recommender.py --input prompt_file.json

5️⃣ Evaluate Results

Description: Computes evaluation metrics using the recommendation results from the previous step.

python DRAGRU/movie-lens/evaluation.py --input recommender_output.json

🤝 Contributing

Contributions, suggestions, and pull requests are welcome. Feel free to ask for improvements (README, visualization, scripts, etc.).

⭐ If You Find This Useful

Please consider ⭐ starring the repository — it's the best way to support this project.

Citation

@article{zhang2025customized,
  title={Customized Retrieval-Augmented Generation with LLM for Debiasing Recommendation Unlearning},
  author={Zhang, Haichao and Zhang, Chong and Hu, Peiyu and Qiu, Shi and Wang, Jia},
  journal={arXiv preprint arXiv:2511.05494},
  year={2025}
}

---