Turbo-CF

Overview and Features

This repo is the source code for Turbo-CF (SIGIR 2024)

• Training-Free: Operates without the need for model training.
• Polynomial Graph Filters: Utilizes efficient polynomial filters, makes it possible to easily compute in parallelism via GPU.
• High Accuracy: Achieves near SOTA with very fast computation

Installation

Clone the repository and install the necessary dependencies:

git clone https://github.com/jindeok/Turbo-CF.git
cd Turbo-CF
pip install -r requirements.txt

Running

To run the Turbo-CF (dataset:'gowalla', 'yelp', 'amazon'):

python main.py --dataset [dataset_name] --filter [filter_type] --alpha [alpha] --power [power]

example:

python main.py --dataset gowalla --filter 1

Can also be run in jupyter env: main.ipynb

Optimal parameters

We provide optimal parameters for each dataset herein.

• Gowalla alpha (a): 0.6, power (s): 0.7 filter: 1
• Yelp alpha (a): 0.6, power (s): 1 filter: 2
• Amazon-book alpha (a): 0.5, power (s): 1.4 filter: 1

(We note that as Turbo-CF is a rule-based method, we use all available dataset (training + validation) for the graph filtering (inference).)

Citation

If this work was helpful for your project, please kindly cite this in your paper

[ref] Jin-Duk Park, Yong-Min Shin, and Won-Yong Shin. "Turbo-CF: Matrix Decomposition-Free Graph Filtering for Fast Recommendation." In SIGIR 2024.

@inproceedings{park2024turbo,
title={Turbo-cf: Matrix decomposition-free graph filtering for fast recommendation},
author={Park, Jin-Duk and Shin, Yong-Min and Shin, Won-Yong},
booktitle={Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages={2672--2676},
year={2024}
}