This is an implementation for our SIGIR 2020 paper: How to Retrain Recommender System? A Sequential Meta-Learning Method.
Contributors: Yang Zhang, Chenxu Wang, Fuli Feng, Xiangnan He
pytorch >= 1.2
numpy
- --MF_lr: learning rate for
$\hat(w)_t$ - --TR_lr: learning rate for Transfer
- --l2：
$\lambda_1$ in paper - --TR_l2:
$\lambda_2$ in paper - --MF_epochs: epochs of learning MF
$\hat(w)_t$ (line 6 in Alg 1 in paper) - --TR_epochs: epochs of learning Transfer
$\theta$ (line 9 in Alg 1 in paper) - --multi_num: stop condition (line 4 in Alg 1 in paper)
- others: read help, or "python main_yelp.py --help"
Save as array,
- train: period_num.npy (user_id,item_id)
- test: period_num.npy (user_id,pos_item_id, neg_item_id)
- Dataset URL: https://rec.ustc.edu.cn/share/1e40a9b0-c80a-11ea-b178-77f793cf5b55 (This is Yelp, Adressa will be uploaded later)
nohup python main_yelp.py --MF_epochs=1 --TR_epochs=1 --multi_num=10 > yelp_log.out &
or
python main_yelp.py --MF_epochs=1 --TR_epochs=1 --multi_num=10
nohup python main_news.py --MF_epochs=2 --TR_epochs=2 --multi_num=7 > yelp_log.out &
for Adressa, if we set the MF_epochs and TR_epochs same to paper (=1) , we can also get a similar performance if we adjust multi_num.