/SimGCL-MixGCF

Primary LanguagePython

Development

LightGCN_1110.py

能够同时mixup多个samples，而且权重自适应分配。

Requirements

numba==0.53.1
numpy==1.20.3
scipy==1.6.2
tensorflow==1.14.0
torch>=1.7.0

Usage

  1. Configure the xx.conf file in the directory named conf. (xx is the name of the model you want to run)
  2. Run main.py and choose the model you want to run.

Implemented Models

Model Paper Type Code
XSimGCL Yu et al. XSimGCL: Towards Extremely Simple Graph Contrastive Learning for Recommendation, Submitted to TKDE. Graph + CL PyTorch
SimGCL Yu et al. Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation, SIGIR'22. Graph + CL PyTorch
DirectAU Wang et al. Towards Representation Alignment and Uniformity in Collaborative Filtering, KDD'22. Graph PyTorch
NCL Lin et al. Improving Graph Collaborative Filtering with Neighborhood-enriched Contrastive Learning, WWW'22. Graph + CL PyTorch
MixGCF Huang et al. MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems, KDD'21. Graph + DA PyTorch
MHCN Yu et al. Self-Supervised Multi-Channel Hypergraph Convolutional Network for Social Recommendation, WWW'21. Graph + CL TensorFlow
SGL Wu et al. Self-supervised Graph Learning for Recommendation, SIGIR'21. Graph + CL TensorFlow & Torch
SEPT Yu et al. Socially-Aware Self-supervised Tri-Training for Recommendation, KDD'21. Graph + CL TensorFlow
BUIR Lee et al. Bootstrapping User and Item Representations for One-Class Collaborative Filtering, SIGIR'21. Graph + DA PyTorch
SSL4Rec Yao et al. Self-supervised Learning for Large-scale Item Recommendations, CIKM'21. Graph + CL PyTorch
SelfCF Zhou et al. SelfCF: A Simple Framework for Self-supervised Collaborative Filtering, arXiv'21. Graph + DA PyTorch
LightGCN He et al. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation, SIGIR'20. Graph PyTorch
MF Yehuda et al. Matrix Factorization Techniques for Recommender Systems, IEEE Computer'09. Graph PyTorch
* CL is short for contrastive learning (including data augmentation); DA is short for data augmentation only

Leaderboard

The results are obtained on the dataset of Yelp2018. We performed grid search for the best hyperparameters.
General hyperparameter settings are: batch_size: 2048, emb_size: 64, learning rate: 0.001, L2 reg: 0.0001.

Model Recall@20 NDCG@20 Hyperparameter settings
MF 0.0543 0.0445
LightGCN 0.0639 0.0525 layer=3
NCL 0.0670 0.0562 layer=3, ssl_reg=1e-6, proto_reg=1e-7, tau=0.05, hyper_layers=1, alpha=1.5, num_clusters=2000
SGL 0.0675 0.0555 λ=0.1, ρ=0.1, tau=0.2 layer=3
MixGCF 0.0691 0.0577 layer=3, n_nes=64, layer=3
DirectAU 0.0695 0.0583 𝛾=2, layer=3
SimGCL 0.0721 0.0601 λ=0.5, eps=0.1, tau=0.2, layer=3
XSimGCL 0.0723 0.0604 λ=0.2, eps=0.2, l∗=1 tau=0.15 layer=3

Implement Your Model

  1. Create a .conf file for your model in the directory named conf.
  2. Make your model inherit the proper base class.
  3. Reimplement the following functions.
    • build(), train(), save(), predict()
  4. Register your model in main.py.

Related Datasets

   
Data Set Basic Meta User Context
Users ItemsRatings (Scale) Density Users Links (Type)
Douban 2,848 39,586 894,887 [1, 5] 0.794% 2,848 35,770 Trust
LastFM 1,892 17,632 92,834 implicit 0.27% 1,892 25,434 Trust
Yelp 19,539 21,266 450,884 implicit 0.11% 19,539 864,157 Trust
Amazon-Book 52,463 91,599 2,984,108 implicit 0.11% - - -