GradPU: Positive-Unlabeled Learning via Gradient Penalty and Positive Upweighting
Authors: Songmin Dai, Xiaoqiang Li, Yue Zhou, Xichen Ye, Tong Liu
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our proposed Grad PU on three datasets: MNIST, Fashion MNIST, and CIFAR10. The results demonstrate that Grad PU achieves state-of-the-art performance on both unbiased and biased positive labeling scenarios. |
| Researcher Affiliation | Academia | Songmin Dai, Xiaoqiang Li*, Yue Zhou, Xichen Ye, Tong Liu, School of Computer Engineering and Science, Shanghai University, China {laodar, xqli, yuez, yexichen0930, tong_liu}@shu.edu.cn |
| Pseudocode | Yes | Algorithm 1: Training procedure of Grad PU classifier. |
| Open Source Code | No | The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | MNIST: The MNIST dataset contains 10 classes of grayscale handwritten digit images with 28 28 pixels. It consists of 60,000 training images and 10,000 test images. As in (Chen et al. 2020), we choose odd numbers 1, 3, 5, 7, and 9 as the positive class while other numbers as the negative class. Fashion MNIST: The Fashion MNIST dataset contains 10 classes of gray-scale fashion product images. Like MNIST, it consists of 60,000 training images and 10,000 test images, with the size of 28 28 pixels. CIFAR10: The CIFAR10 dataset consists of colored images with sizes of 32 32 pixels. There are 50,000 training images and 10,000 test images. |
| Dataset Splits | Yes | We follow the protocol widely used by randomly selecting a small portion of positive instances from the original training data as labeled positive Xp, 500 instances as the validation set Xval, and the remaining instances as unlabeled data Xu. |
| Hardware Specification | No | The paper mentions 'The High Performance Computing Center of Shanghai University, and Shanghai Engineering Research Center of Intelligent Computing System for providing the computing resources and technical support,' but it does not specify any exact GPU/CPU models, processor types, or memory amounts used for the experiments. |
| Software Dependencies | No | The paper mentions using the 'Adam optimizer with β1 = 0.9 and β2 = 0.999' but does not specify its version number, nor does it list any other software or library dependencies with their specific version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We use the Adam optimizer with β1 = 0.9 and β2 = 0.999. For all trials, we use the batch size of 250 and gradually decrease the learning rate to 0.1 times of its initial value. We report the mean accuracies and deviations over three runs. For MNIST and Fashion MNIST, we train the models for 100 epochs, while for CIFAR10, we train the models for 200 epochs. We tune hyper-parameters such as weight decay, learning rate, the gradient penalty strength α, and the upweighting strength βmax using a validation set. |