This is the Pytorch implementation for reproducing the results in
CIKM 2021. Yifei Shen, Yongji Wu, Yao Zhang, Caihua Shan, Jun Zhang, Khaled B. Letaief, Dongsheng Li(2021). How Powerful is Graph Convolution for Recommendation? Paper in arXiv
The code is heavily built on LightGCN's code.
SIGIR 2020. Xiangnan He, Kuan Deng ,Xiang Wang, Yan Li, Yongdong Zhang, Meng Wang(2020). LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation, Paper in arXiv.
(See Pytorch implementation) We also adopt exactly the same dataset and train/test splitting. The code is not optimized for speed but rather for simplicity.
pip install -r requirements.txt
run untrained LightGCN on Gowalla dataset:
- change base directory
cd Fig1
Change ROOT_PATH in code/world.py
- command
./run.sh
- log output
run untrained LGCN-IDE or GF-CF or LightGCN on Amazon-book dataset:
- change base directory
cd Table3\&4
Change ROOT_PATH in code/world.py
- To reproduce the results for LightGCN in Table 3&4
python3 main.py --decay=1e-4 --lr=0.001 --layer=3 --seed=2020 --dataset="amazon-book" --topks="[20]" --recdim=256
- To reproduce the results for LGCN-IDE in Table 3&4
python3 main.py --dataset="amazon-book" --topks="[20]" --simple_model "lgn-ide"
- To reproduce the results for GF-CF in Table 3&4
python3 main.py --dataset="amazon-book" --topks="[20]" --simple_model "gf-cf"
- log output