MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems

This is our PyTorch implementation for the paper:

Tinglin Huang, Yuxiao Dong, Ming Ding, Zhen Yang, Wenzheng Feng, Xinyu Wang, Jie Tang (2021). MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems. Paper link. In KDD'2021, Virtual Event, Singapore, August 14-18, 2021.

Author: Mr. Tinglin Huang (tinglin.huang at zju.edu.cn)

Citation

If you want to use our codes in your research, please cite: ​

@inproceedings{MixGCF2021,
  author    = {Tinglin Huang and
               Yuxiao Dong and
               Ming Ding and
               Zhen Yang and
               Wenzheng Feng and
               Xinyu Wang and
               Jie Tang},
  title     = {MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems},
  booktitle = {{KDD}},
  year      = {2021}
}

Environment Requirement

The code has been tested running under Python 3.7.6. The required packages are as follows:

• pytorch == 1.7.0
• numpy == 1.20.2
• scipy == 1.6.3
• sklearn == 0.24.1
• prettytable == 2.1.0

Training

The instruction of commands has been clearly stated in the codes (see the parser function in utils/parser.py). Important argument:

• K
  • It specifies the number of negative instances in K-pair loss. Note that when K=1 (by default), the K-pair loss will degenerate into the BPR pairwise loss.
• n_negs
  • It specifies the size of negative candidate set when using MixGCF.
• ns
  • It indicates the type of negative sample method. Here we provide two options: rns and mixgcf.

LightGCN

Random sample(rns)

python main.py --dataset ali --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns rns --K 1 --n_negs 1

python main.py --dataset yelp2018 --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns rns --K 1 --n_negs 1

python main.py --dataset amazon --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns rns --K 1 --n_negs 1

MixGCF

python main.py --dataset ali --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns mixgcf --K 1 --n_negs 32

python main.py --dataset yelp2018 --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns mixgcf --K 1 --n_negs 64

python main.py --dataset amazon --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 3 --pool mean --ns mixgcf --K 1 --n_negs 16

NGCF

Random sample(rns)

python main.py --dataset ali --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns rns --K 1 --n_negs 1

python main.py --dataset yelp2018 --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns rns --K 1 --n_negs 1

python main.py --dataset amazon --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns rns --K 1 --n_negs 1

MixGCF

python main.py --dataset ali --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns mixgcf --K 1 --n_negs 64

python main.py --dataset yelp2018 --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns mixgcf --K 1 --n_negs 64

python main.py --dataset amazon --gnn ngcf --dim 64 --lr 0.0001 --batch_size 1024 --gpu_id 0 --context_hops 3 --pool concat --ns mixgcf --K 1 --n_negs 64

The training log is also provided. The results fluctuate slightly under different running environment.

Dataset

We use three processed datasets: Alibaba, Yelp2018, and Amazon.

	Alibaba	Yelp2018	Amazon
#Users	106,042	31,668	192,403
#Items	53,591	38,048	63,001
#Interactions	907,407	1,561,406	1,689,188
Density	0.00016	0.00130	0.00014