DPL: Debiased Pairwise Learning for Implicit Collaborative Filtering

This work is accepted for publication in IEEE Transactions on Knowledge and Data Engineering (IEEE TKDE). Should you use this work in your research, please cite the following paper:

# bibtex
@article{liu2024DPL,
  title={Debiased pairwise learning for implicit collaborative filtering},
  author={Liu, Bin and Luo, Qin and Wang, Bang},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2024},
  volume={36},
  number={12},
  pages={7878-7892},
  doi = {10.1109/TKDE.2024.3479240}
}

DPL rectifys the probability bias caused by sampling bias, in order to obtain an estimator consistent with the supervised paired loss. This correction of optimization objective bias enables collaborative filtering models to learn parameters similar to those in fully supervised settings, thereby achieving improved generalization performance.

Prerequisites

• Python 3.7
• PyTorch 1.3.1

Flags:

--LOSS: loss function, choose DPL or BPR.

--dataset: dataset name, choose 100k, 1M, gowalla or yelp2018.

--tau_plus: positive class prior.

--l2: l2 regularization constant.

--lr: learning rate.

--encoder : backbones, choose MF or LightGCN.

--M: Additional positive examples for each user, $M \geq 1$.

--N: Negative examples for each user, $N \geq 1$.

Model Pretraining

For instance, run the following command to train an embedding on different datasets.

python main.py  --LOSS DPL  --dataset_name '100k'  --encoder MF  --tau_plus 0.07  --lr 1e-3 --l2 1e-5 

python main.py  --LOSS DPL  --dataset_name '1M'  --encoder MF  --tau_plus 0.05  --lr 1e-3 --l2 1e-6

Parameter Settings

Dataset	Backbone	$\tau^+$	Learning Rate	l2	Batch Size	$M$	$N$
MovieLens 100K	MF	7e-2	1e-3	1e-5	1024	10	10
MovieLens 100K	LightGCN	7e-2	1e-3	1e-5	1024	5	1
MovieLens 1M	MF	5e-2	1e-3	1e-6	1024	10	10
MovieLens 1M	LightGCN	5e-2	1e-3	1e-6	1024	5	1
Yelp2018	MF	1e-2	1e-3	1e-6	1024	5	5
Yelp2018	LightGCN	1e-2	1e-3	1e-6	1024	10	1
Gowalla	MF	1e-2	1e-3	1e-6	1024	5	5
Gowalla	LightGCN	1e-2	1e-3	1e-6	1024	5	1

Acknowledgements

We extend our appreciation to DCL that inspired us to advance research on debiasing pairwise loss.