A Variational Approach for Learning from Positive and Unlabeled Data
Authors: Hui Chen, Fangqing Liu, Yin Wang, Liyue Zhao, Hao Wu
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We illustrate the effectiveness of the proposed variational method on a number of benchmark examples.Our theoretical and experimental analysis demonstrates that, in contrast with the existing methods, the variational principle based method can achieve high classification accuracies in PU learning tasks without the estimation of class prior or the assumption of data separability. |
| Researcher Affiliation | Collaboration | Hui Chen School of Mathematical Science Tongji University, Shanghai, P. R. China...Liyue Zhao Cloudwalk Inc. Shanghai, P. R. China |
| Pseudocode | Yes | A stochastic gradient based implementation of VPU with loss function defined by (7) and (8) is given in Algorithm 1 |
| Open Source Code | Yes | All the other detailed settings of datasets and algorithms are provided in Section B of Suppl. Material, and the software code for VPU is also available5. 5https://github.com/HC-Feynman/vpu |
| Open Datasets | Yes | We conduct experiments on three benchmark datasets taken from the UCI Machine Learning Repository [34, 35], and the classification results are reported in Table 2.Here we compare all the methods on three image datasets: Fashion MNIST, CIFAR-10, and STL-10.Datasets are downloaded from https://github.com/zalandoresearch/fashion-mnist, https://www.cs.toronto.edu/~kriz/cifar.html and http://cs.stanford.edu/~acoates/stl10. |
| Dataset Splits | Yes | In all experiments, α is chosen as 0.3 and λ {1e 4, 3e 4, 1e 3, , 1, 3} is determined by holdout validation unless otherwise specified. By default, the accuracies are evaluated on test sets and the mean and standard deviation values are computed from 10 independent runs. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., CPU/GPU models, memory) used for running the experiments. |
| Software Dependencies | Yes | Notice that in the rest of the paper, we denote the 10 classes of each image datasets with integers ranging from 0 to 9, following the default settings in torchvision 0.5.0 |
| Experiment Setup | Yes | In all experiments, α is chosen as 0.3 and λ {1e 4, 3e 4, 1e 3, , 1, 3} is determined by holdout validation unless otherwise specified. We use Adam as the optimizer for VPU with hyperparameters (β1, β2) = (0.5, 0.99).the classifiers (including discriminators of Gen PU) are modeled by 7-layer MLP for UCI datasets, Le Net-5 [33] for Fashion MNIST and 7-layer CNN for CIFAR-10 and STL-10. |