AdvInfoNCE

Overview

Code of "Empowering Collaborative Filtering Generalization via Principled Adversarial Contrastive Loss"

Run the Code

• We provide implementation for various baselines presented in the paper.

• To run the code, first run the following command to install tools used in evaluation:

python setup.py build_ext --inplace

LightGCN backbone

• AdvInfoNCE Training:

Tencent:

python train_AdvInfoNCE.py --train_norm --pred_norm --modeltype AdvInfoNCE --model_version embed --dataset tencent_synthetic --n_layers 2 --lr 1e-3 --batch_size 2048 --neg_sample 128 --tau 0.09 --adv_lr 5e-5 --eta_epochs 7 --k_neg 64 --patience 20 --dsc sota_tencent_lgn

KuaiRec:

python train_AdvInfoNCE.py --train_norm --pred_norm --modeltype AdvInfoNCE --model_version embed  --dataset kuairec2 --n_layers 2 --neg_sample 128 --tau 2 --lr 3e-5 --batch_size 2048 --adv_lr 5e-5 --eta_epochs 12 --warm_up_epochs 0 --adv_interval 5 --dsc sota_kuairec_lgn

Yahoo:

python train_AdvInfoNCE.py --train_norm --pred_norm --model_version embed --modeltype AdvInfoNCE --dataset yahoo.new --n_layers 2 --neg_sample 64 --tau 0.28 --lr 5e-4 --batch_size 1024 --adv_lr 1e-4 --eta_epochs 13 --dsc sota_yahoo_lgn

Coat:

python train_AdvInfoNCE.py --train_norm --pred_norm --modeltype AdvInfoNCE --model_version embed --dataset coat --n_layers 2 --neg_sample 64 --tau 0.75 --lr 1e-3 --batch_size 1024 --adv_lr 1e-2 --eta_epochs 20 --adv_interval 15 --dsc sota_coat_lgn

• INFONCE Training:

python main.py --train_norm --pred_norm --modeltype  INFONCE --dataset kuairec2 --n_layers 2 --batch_size 2048 --lr 3e-5 --neg_sample 128 --tau 2  --dsc infonce

MF backbone

• AdvInfoNCE Training:

Tencent:

python train_AdvInfoNCE.py --train_norm --pred_norm --modeltype AdvInfoNCE --model_version embed --dataset tencent_synthetic --n_layers 0 --lr 1e-3 --batch_size 2048 --neg_sample 128 --tau 0.09 --adv_lr 5e-5 --eta_epochs 8 --patience 20 --dsc sota_tencent

Yahoo:

python train_AdvInfoNCE.py --train_norm --pred_norm --model_version embed --modeltype AdvInfoNCE --dataset yahoo.new --n_layers 0 --neg_sample 64 --tau 0.28 --lr 5e-4 --batch_size 1024 --adv_lr 1e-4 --eta_epochs 12 --dsc sota_yahoo

Coat:

python train_AdvInfoNCE.py --train_norm --pred_norm --modeltype AdvInfoNCE --model_version embed --dataset coat --n_layers 0 --neg_sample 64 --tau 0.75 --lr 1e-3 --batch_size 1024 --adv_lr 1e-2 --eta_epochs 18 --adv_interval 15 --dsc sota_coat

• InfoNCE Training:

python main.py --train_norm --pred_norm --modeltype INFONCE --dataset tencent_synthetic --n_layers 0 --tau 0.09 --neg_sample 128 --batch_size 2048 --lr 1e-3 --dsc infonce

Requirements

• python == 3.7.10

• pytorch == 1.12.1+cu102

• tensorflow == 1.14

• reckit == 0.2.4

Reference

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

@inproceedings{AdvInfoNCE,
  title={Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss},
  author={Zhang, An and Sheng, Leheng and Cai, Zhibo and Wang, Xiang and Chua, Tat-Seng},
  booktitle={{NeurIPS}},
  year={2023}
}