Python implementation for our paper "Efficient-FedRec: Efficient Federated Learning Frameworkfor Privacy-Preserving News Recommendation" in EMNLP 2021.
Directly applying federated learning on news recommendation models will lead to high computation and communication cost on user side. In this work, we propose Efficient-FedRec, in which we dicompose the news recommendation model into a large news model maintained on server and a light-weight user model computed on the user side. Experiments on two public dataset show the effectiveness of our method.
Requirments
numpy
torch==1.9.1
transformers==4.12.5
tqdm
sklearn
wandb
- Download datasets
cd raw
chmod +x download.sh
./download.sh mind .
./download.sh adressa .
- Preprocess datasets
cd preprocess
# modify adressa to mind format
python adressa_raw.py
# preprocess mind dataset
python news_process.py --data mind
python user_process.py --data mind
# preprocess adressa dataset
python news_process.py --data adressa
python user_process.py --data adressa
- Run experiments
# You may need to configure your wandb account first
cd src
python main.py --data mind
# get prediction result of the best checkpoint and submit on condalab
python main.py --data mind --mode predict
# train on adressa
python main.py --data adressa --max_train_steps 500 --validation_step 10 --bert_type NbAiLab/nb-bert-base
# test on adressa
python main.py --data adressa --mode test --bert_type NbAiLab/nb-bert-base
Wandb result on MIND dataset
Zip the prediction.txt file and upload to MIND competition. Test result is
Wandb result on Adressa dataset.
Test result is
test auc: 0.7980, mrr: 0.4637, ndcg5: 0.4852, ndcg10: 0.5497
If you want to cite Efficient-Fedrec in your papers (much appreciated!), you can cite it as follows:
@inproceedings{yi-etal-2021-efficient,
title = "Efficient-{F}ed{R}ec: Efficient Federated Learning Framework for Privacy-Preserving News Recommendation",
author = "Yi, Jingwei and
Wu, Fangzhao and
Wu, Chuhan and
Liu, Ruixuan and
Sun, Guangzhong and
Xie, Xing",
booktitle = "EMNLP",
year = "2021",
pages = "2814--2824"
}