Positive Distribution Pollution: Rethinking Positive Unlabeled Learning from a Unified Perspective
Authors: Qianqiao Liang, Mengying Zhu, Yan Wang, Xiuyuan Wang, Wanjia Zhao, Mengyuan Yang, Hua Wei, Bing Han, Xiaolin Zheng
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the superiority of Co VPU over the state-of-the-art PU learning methods under these problems. and In this section, we present the extensive experiments to answer the following question: Q1: How does Co VPU perform when compared to the state-of-the-art PU learning methods? Q2: How do Co VPU s key components contribute to its performance? Q3: How does Co VPU perform in addressing the three challenging problems? |
| Researcher Affiliation | Collaboration | 1 College of Computer Science, Zhejiang University, China 2 School of Computing, Macquarie University, Australia 3 MYbank, Ant Group, China |
| Pseudocode | Yes | Algorithm 1: Training process of Co VPU |
| Open Source Code | No | The paper does not explicitly state that source code for the methodology is provided or available. |
| Open Datasets | Yes | In experiments, we adopt two widely used PU learning datasets, i.e., MNIST and Cifar10 (Kiryo et al. 2017; Su, Chen, and Xu 2021; Hu et al. 2021). |
| Dataset Splits | Yes | For data preprocessing, we chronologically split the time series in Bank Loan with ratio 0.6/0.1/0.3 as train/validation/test dataset, and adopt the train/validation/test dataset splits as provided by the original datasets for MNIST and Cifar10, which have been widely adopted in PU leering studies (Kiryo et al. 2017; Su, Chen, and Xu 2021; Chen et al. 2020a). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' but does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | In our proposed Co VPU, we set π to be 0.5 and J to be 20, and we tune all other hyperparameters through grid search. We also carefully tune the hyper-parameters of all comparison methods through cross-validation to achieve their best performance. We provide the ground-truth class prior values to all the comparison methods that require the input of class prior. For training details, we adopt Adam optimizer (Kingma and Ba 2015) and tune the learning rate and weight decay by a grid select in [10 2, 10 5] for all methods. We set the batch size to 1024 for batch training. |