GRecX

An Efficient and Unified Benchmark for GNN-based Recommendation.

Homepage and Documentation

• Homepage: https://github.com/maenzhier/GRecX
• Paper: GRecX: An Efficient and Unified Benchmark for GNN-based Recommendation

Example Benchmark: Performance on Yelp and Gowalla with BPR Loss

Performance on Yelp with BPR Loss:

Performance on Gowalla with BPR Loss:

Demo

We recommend you get started with some demos.

• Matrix Factorization (MF)
• MLP + MF
• NGCF
• LightGCN
• UltraGCN

Preliminary Comparison

LightGCN-Yelp dataset (featureless)

• BCE-loss

Algo	Precision@10	Precision@20	Recall@10	Recall@20	nDCG@10	nDCG@20
MF	0.029597	0.025495	0.032733	0.056086	0.037332	0.045805
NGCF	0.024713	0.021893	0.028251	0.049611	0.031357	0.039549
LightGCN	---	---	---	---	0.037350	0.045872
UltraGCN-single	0.030652	0.026790	0.033913	0.058886	0.038576	0.047766
UltraGCN	0.03553	0.030346	0.039526	0.067028	0.045365	0.055376

• BPR-loss

Algo	Precision@10	Precision@20	Recall@10	Recall@20	nDCG@10	nDCG@20
MF	0.031489	0.027303	0.034733	0.060333	0.040103	0.049406
NGCF	0.030375	0.026699	0.034502	0.059984	0.038732	0.048351
LightGCN	0.033544	0.028996	0.037277	0.064128	0.042907	0.052667
UltraGCN-single	---	---	---	---	---	---
UltraGCN	---	---	---	---	---	---

Note that "UltraGCN-single" uses loss with one negative sample and one negatvie loss weight

Cite

If you use GRecX in a scientific publication, we would appreciate citations to the following paper:

@misc{cai2021grecx,
title={GRecX: An Efficient and Unified Benchmark for GNN-based Recommendation},
author={Desheng Cai and Jun Hu and Shengsheng Qian and Quan Fang and Quan Zhao and Changsheng Xu},
year={2021},
eprint={2111.10342},
archivePrefix={arXiv},
primaryClass={cs.IR}
}