A Variational Approach for Learning from Positive and Unlabeled Data

This repository is the official implementation of A Variational Approach for Learning from Positive and Unlabeled Data.

Requirements

To install requirements:

pip install -r requirements.txt

Training and Evaluation

To repeat experiments in the paper, run the following commands:

python run.py --dataset cifar10 --lam 0.03 --num_labeled 3000 --learning-rate 3e-5 --gpu <gpu_id>

python run.py --dataset fashionMNIST --lam 0.3 --num_labeled 3000 --learning-rate 3e-4 --gpu <gpu_id>

python run.py --dataset stl10 --lam 0.3 --num_labeled 2250 --learning-rate 1e-4 --gpu <gpu_id>

python run.py --dataset pageblocks --lam 0.0001 --num_labeled 100 --learning-rate 3e-4 --batch-size 100 --gpu <gpu_id>

python run.py --dataset grid --lam 0.1 --num_labeled 1000 --learning-rate 3e-4 --gpu <gpu_id>

python run.py --dataset avila --lam 0.1 --num_labeled 2000 --learning-rate 6e-4 --gpu <gpu_id>

Results

Our model achieves the following performance (accuracy) on PU learing tasks of FashionMNIST, CIFAR-10 and STL-10:

Model	FashionMNIST	CIFAR-10	STL-10
VPU	92.7%	89.5%	79.7%
nnPU	90.8%	85.6%	78.3%

where nnPU is the current state-of-the-art. For more details, please refer to Table 2 and 3 in the paper.

Cite the paper

If you find this useful, please cite

@article{chen2019variational,
  title={A Variational Approach for Learning from Positive and Unlabeled Data},
  author={Chen, Hui and Liu, Fangqing and Wang, Yin and Zhao, Liyue and Wu, Hao},
  journal={arXiv preprint arXiv:1906.00642},
  year={2019}
}