nnPUSB

Learning from positive and unlabeled data with a selection bias(nnPUSB) reproductive code on MNIST

Some code comes from kiryor's.

Details

Requirements

Python 3
Pytorch >= 1.0
Scikit-learn >= 0.2
Numpy >=1.1

Quick start

You can set your GPU on the global variable device in args.py

You can run an example code of MNIST for comparing the performance of nnPU learning and nnPUSB learning.

python3 train.py

You can see additional information by adding --help.

Example result

nnPU / nnPUSB Precision Recall in result/precision_recall.png

Reference

[1] Masahiro Kato and Takeshi Teshima and Junya Honda. "Learning from Positive and Unlabeled Data with a Selection Bias." International Conference on Learning Representations. 2019.

[2] Ryuichi Kiryo, Gang Niu, Marthinus Christoffel du Plessis, and Masashi Sugiyama. "Positive-Unlabeled Learning with Non-Negative Risk Estimator." Advances in neural information processing systems. 2017.