nips23_social_igf

This is the implementation of our paper accepted by NeurIPS' 23:

Fairly Recommending with Social Attributes: A Flexible and Controllable Optimization Approach

• Jinqiu Jin, Haoxuan Li, Fuli Feng, Sihao Ding, Peng Wu, Xiangnan He

To reproduce the results of our proposed method, there are two steps:

• Step 1: Pretrain a BPRMF model

  To initialize all the baselines and ablations with a well-trained BPRMF model, you can run the following commands:

  python bprmf.py --dataset KuaiRec --batch_size 4096 --lr 1e-3 --reg 1e-5 --max_epoch 200 --patience 10
  python bprmf.py --dataset Epinions --batch_size 8192 --lr 1e-2 --reg 1e-5 --max_epoch 200 --patience 10

  This step is optional, since we have provided the well-trained bprmf model in ./data/dataset/best_model.pth.tar.

• Step 2: Run our proposed method

  run python main.py --dataset KuaiRec --batch_size 4096 --reg 1e-5 --max_epoch 50 --patience 5 (for KuaiRec) or python main.py --dataset Epinions --batch_size 8192 --reg 1e-5 --max_epoch 50 --patience 5 (for Epinions) plus the following arguments:

  --argument	description	(optimal) values
  --pref_idx	the preference (solution) region	0, 1, 2, 3, 4
  --lr	lr for solving the subproblem	see the parameter config page
  --initsol_lr	lr for finding the initial solution	swee the parameter config page
  --thre	accuracy threhold $\xi$ in Eq.(11)	0.34 - KuaiRec, 0.405 - Epinions
  --mode	trade-off between SP/NSP or EO/NEO	sp, eo

• Below are hyper-parameters for reproducing the main results:

  • For KuaiRec dataset: run python main.py --dataset KuaiRec --batch_size 4096 --reg 1e-5 --max_epoch 50 --patience 5 plus the following arguments:

    --pref_idx 0 --lr 1e-4 --initsol_lr 3e-4 --thre 0.34 --mode sp
    --pref_idx 1 --lr 1e-4 --initsol_lr 1e-3 --thre 0.34 --mode sp
    --pref_idx 2 --lr 3e-3 --initsol_lr 3e-3 --thre 0.34 --mode sp
    --pref_idx 3 --lr 3e-4 --initsol_lr 3e-4 --thre 0.34 --mode sp
    --pref_idx 4 --lr 1e-3 --initsol_lr 3e-4 --thre 0.34 --mode sp
    
    --pref_idx 0 --lr 3e-5 --initsol_lr 1e-3 --thre 0.4 --mode eo
    --pref_idx 1 --lr 1e-4 --initsol_lr 3e-4 --thre 0.5 --mode eo
    --pref_idx 2 --lr 1e-4 --initsol_lr 3e-3 --thre 0.34 --mode eo
    --pref_idx 3 --lr 1e-4 --initsol_lr 3e-3 --thre 0.4 --mode eo
    --pref_idx 4 --lr 3e-4 --initsol_lr 3e-3 --thre 0.34 --mode eo

  • For Epinions dataset: run python main.py --dataset Epinions --batch_size 8192 --reg 1e-5 --max_epoch 50 --patience 5 plus the following options:

    --pref_idx 0 --lr 3e-4 --initsol_lr 3e-3 --thre 0.6 --mode sp
    --pref_idx 1 --lr 1e-3 --initsol_lr 1e-2 --thre 0.405 --mode sp
    --pref_idx 2 --lr 1e-3 --initsol_lr 1e-2 --thre 0.405 --mode sp
    --pref_idx 3 --lr 1e-2 --initsol_lr 1e-2 --thre 0.405 --mode sp
    --pref_idx 4 --lr 3e-3 --initsol_lr 3e-2 --thre 0.5 --mode sp
    
    --pref_idx 0 --lr 3e-4 --initsol_lr 3e-2 --thre 0.405 --mode eo
    --pref_idx 1 --lr 1e-3 --initsol_lr 3e-2 --thre 0.405 --mode eo
    --pref_idx 2 --lr 1e-2 --initsol_lr 1e-2 --thre 0.405 --mode eo
    --pref_idx 3 --lr 3e-3 --initsol_lr 1e-2 --thre 0.405 --mode eo
    --pref_idx 4 --lr 1e-3 --initsol_lr 3e-2 --thre 0.405 --mode eo

• Besides, we also provide the code of other baselines and ablations. The following table shows the usage of these codes:

  File	Method	Usage (Example)
  regularization.py	SP(EO) Reg, NSP(NEO) Reg, SP&NSP(EO&NEO) Reg	python regularization.py --lr 1e-4 --reg1 2 --mode eo --dataset KuaiRec
  post_reconstruct.py	SP(EO) Post, NSP(NEO) Post, SP&NSP(EO&NEO) Post	python post_reconstruct.py --lr 3e-5 --reg1 0.1 --reg2 0.1 --mode sp_nsp --dataset KuaiRec
  moomtl.py	MOOMTL	python moomtl.py --lr 1e-3 --initsol_lr 1e-3 --dataset Epinions
  wo_con.py	w/o con	python wo_con.py --lr 3e-4 --mode sp --initsol_lr 1e-2 --pref_idx 0 --dataset Epinions
  single_obj.py	w/o NSP(NEO), w/o SP(EO)	python single_obj.py --lr 3e-3 --mode nsp --thre 0.4 --dataset KuaiRec
  wo_region.py	w/o region	python wo_region.py --lr 1e-3 --thre 0.405 --mode eo --dataset Epinions

• We also provide the hyper-parameters for reproducing all results of various baselines and abalations here.

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

  @inproceedings{social_igf,   
        author    = {Jinqiu Jin and
                     Haoxuan Li and 
                     Fuli Feng and 
                     Sihao Ding and
                     Peng Wu and
                     Xiangnan He}, 
        title     = {Fairly Recommending with Social Attributes: A Flexible and Controllable Optimization Approach},  
        booktitle = {{NeurIPS}},  
        year      = {2023},   
  }