MACR

This is an implemention for our SIGKDD 2021 paper based on tensorflow

Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System

by Tianxin Wei, Fuli Feng, Jiawei Chen, Ziwei Wu, Jinfeng Yi and Xiangnan He.

Introduction

MACR is a general popularity debias framework based on causal inference and counterfactual reasoning.

Requirements

• tensorflow == 1.14
• Numpy == 1.16.0
• python == 3.6
• Cython == 0.29.24 (Optional)
• CUDA v10

For LightGCN C++ evaluation, please install Cython and do

cd macr_lightgcn
python setup.py build_ext --inplace

Run the code

Like what we mentioned in our paper, great performance can be obtained by easily setting C=30 or C=40. Here we show the example results of setting C=40.

Normal MF :

Change the dataset argument to run experiments on different datasets

python ./macr_mf/train.py --dataset addressa --batch_size 1024 --cuda 0 --saveID 1 --log_interval 10 --lr 0.001 --train normalbce --test normal

MACR MF:

ML10M

python ./macr_mf/train.py --dataset ml_10m --batch_size 8192 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 40 --train rubibceboth --test rubi --alpha 1e-3 --beta 1e-3

Gowalla

python ./macr_mf/train.py --dataset gowalla --batch_size 4096 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 40 --train rubibceboth --test rubi --alpha 1e-2 --beta 1e-3

Globe

python ./macr_mf/train.py --dataset globe --batch_size 4096 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 40 --train rubibceboth --test rubi --alpha 1e-3 --beta 1e-3

Yelp2018

python ./macr_mf/train.py --dataset yelp2018 --batch_size 4096 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 40 --train rubibceboth --test rubi --alpha 1e-2 --beta 1e-3

Adressa

python ./macr_mf/train.py --dataset addressa --batch_size 1024 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 40 --train rubibceboth --test rubi --alpha 1e-3 --beta 1e-3

Normal LightGCN:

python macr_lightgcn/LightGCN.py --data_path data/ --dataset addressa --verbose 1 --layer_size [64,64] --Ks [20] --loss bce --test normal --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 1024 --gpu_id 1 --log_interval 10

MACR LightGCN:

ML10M

python macr_lightgcn/LightGCN.py --data_path data/ --dataset ml_10m --verbose 1 --layer_size [64,64] --Ks [20] --loss bceboth --test rubiboth --c 40 --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 8192 --gpu_id 0 --log_interval 10 --alpha 1e-2 --beta 1e-3

Gowalla

python macr_lightgcn/LightGCN.py --data_path data/ --dataset gowalla --verbose 1 --layer_size [64,64] --Ks [20] --loss bceboth --test rubiboth --c 40 --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 4096 --gpu_id 0 --log_interval 10 --alpha 1e-2 --beta 1e-3

Globe

python macr_lightgcn/LightGCN.py --data_path data/ --dataset globe --verbose 1 --layer_size [64,64] --Ks [20] --loss bceboth --test rubiboth --c 40 --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 4096 --gpu_id 0 --log_interval 10 --alpha 1e-2 --beta 1e-3

Yelp2018

python macr_lightgcn/LightGCN.py --data_path data/ --dataset yelp2018 --verbose 1 --layer_size [64,64] --Ks [20] --loss bceboth --test rubiboth --c 40 --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 4096 --gpu_id 0 --log_interval 10 --alpha 1e-2 --beta 1e-3

Adressa

python macr_lightgcn/LightGCN.py --data_path data/ --dataset addressa --verbose 1 --layer_size [64,64] --Ks [20] --loss bceboth --test rubiboth --c 40 --epoch 2000 --early_stop 1 --lr 0.001 --batch_size 1024 --gpu_id 0 --log_interval 10 --alpha 1e-2 --beta 1e-3

	HR	Rec	NDCG
LightGCN_ML10M	0.15262	0.04745	0.02844
LightGCN_Gowalla	0.28410	0.09076	0.05999
LightGCN_Globe	0.12312	0.05271	0.02854
LightGCN_Adressa	0.16356	0.12967	0.06071
LightGCN_Yelp	0.17210	0.04016	0.02649
MF_ML10M	0.14011	0.04087	0.02407
MF_Gowalla	0.26669	0.08488	0.05756
Mf Globe	0.10725	0.04594	0.02513
MF Adressa	0.13561	0.10612	0.04667
MF Yelp	0.14444	0.02863	0.02039

Fixing C to 40 can get great performance. Tuning the value of C through validation (create your validation set) will get higher performance. For example on ML10M dataset:

# Finetune c
python ./macr_mf/tune.py --dataset ml_10m/val --batch_size 8192 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --start 30 --end 40 --step 11 --train rubibceboth --test rubi --alpha 1e-3 --beta 1e-3 --valid_set valid
# Run experiment
python ./macr_mf/train.py --dataset ml_10m --batch_size 8192 --cuda 0 --saveID 0 --log_interval 10 --lr 0.001 --check_c 1 --c 32 --train rubibceboth --test rubi --alpha 1e-3 --beta 1e-3 --valid_set test

	HR	Rec	NDCG
MF_ML10M (C=40)	0.14011	0.04087	0.02407
MF_ML10M (Validation C=32)	0.14545	0.04235	0.02518

Acknowledgement

Very thanks for Chufeng Shi for his help on code and the LightGCN code repo.

Citation

If you find this paper helpful, please cite our paper.

@inproceedings{wei2021model,
  title={Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System},
  author={Wei, Tianxin and Feng, Fuli and Chen, Jiawei and Wu, Ziwei and Yi, Jinfeng and He, Xiangnan},
  booktitle={Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery \& Data Mining},
  pages={1791--1800},
  year={2021}
}