FairDDA

This repository contains source codes and datasets for our CIKM'25 paper:

• Improving Recommendation Fairness via Graph Structure and Representation Augmentation

Usage

Pre-Train

FairDDA uses pre-trained LightGCN embeddings as input. The pre-trained model parameters are available from the FairMI repository.

Train & Test

• Training FairDDA on MovieLens:

python main.py --dataset=ml-1m

• Training FairDDA on LastFM:

python main.py --dataset=lastfm-360k --batch_size=4096 --ub_reg=40

• Training FairDDA on IJCAI15:

python main.py --dataset=ijcai --num_epochs=70 --sigma=0.5 --reconn_reg=0.1 --lb_reg=0.05 --ub_reg=100

Additional Results

Due to space limitations, the ACM version only analyzes and reports recommendation performance under gender fairness (binary attribute scenario). For additional experiments and analysis in multi-class sensitive scenarios, please refer to Appendix A of our arXiv version.

Citation

If you find this work useful, please cite our paper:

@inproceedings{FairDDA,
    title = {Improving Recommendation Fairness via Graph Structure and Representation Augmentation},
    author = {Xu, Tongxin and Liu, Wenqiang and Bin, Chenzhong and Xiao, Cihan and Zeng, Zhixin and Gu, Tianlong},
    booktitle = {Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
    year = {2025}
}