PULNS: Positive-Unlabeled Learning with Effective Negative Sample Selector
Authors: Chuan Luo, Pu Zhao, Chen Chen, Bo Qiao, Chao Du, Hongyu Zhang, Wei Wu, Shaowei Cai, Bing He, Saravanakumar Rajmohan, Qingwei Lin8784-8792
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental studies on 7 real-world application benchmarks demonstrate that PULNS consistently outperforms the current state-of-the-art methods in PU learning, and our experimental results also confirm the effectiveness of the negative sample selector underlying PULNS. |
| Researcher Affiliation | Collaboration | 1Microsoft Research, China 2Microsoft 365, United States 3The University of Newcastle, Australia 4L3S Research Center, Leibniz University Hannover, Germany 5State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, China 6School of Computer Science and Technology, University of Chinese Academy of Sciences, China |
| Pseudocode | Yes | Algorithm 1: End-to-End Training Process of PULNS |
| Open Source Code | No | The paper mentions that the source codes for competitors (u PU, nn PU, PUSB, PUb N) are available online, but it does not provide an explicit statement or link for the source code of their proposed PULNS method. |
| Open Datasets | Yes | In the context of PU learning, seven benchmarks are commonly used to evaluate PU learning approaches (Kato, Teshima, and Honda 2019): 1) CIFAR-103 and 2) six benchmarks collected from UCI4,5, including MNIST, mushrooms, shuttle, spambase, usps and landsat. Following the common practice, we adopt those seven benchmarks to evaluate the performance of PULNS and its competitors. |
| Dataset Splits | Yes | For benchmarks mushrooms, shuttle, spambase, usps and landsat, the total numbers of samples in the validation set and the testing set are set to 100 and 1000, respectively. For benchmarks CIFAR-10 and MNIST, the total numbers of samples in the validation set and the testing set are set to 500 and 5000, respectively. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions using a convolutional neural network for CIFAR-10 and a multilayer perceptron (MLP) for other benchmarks, and the REINFORCE algorithm for optimization. However, it does not specify any software versions for libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used. |
| Experiment Setup | Yes | The values of |P| and |U| are adopted from the setup used by (Kato, Teshima, and Honda 2019): for benchmarks mushrooms, shuttle, spambase, usps and landsat, |P| and |U| are fixed at 400 and 800, respectively. For benchmarks CIFAR-10 and MNIST, |P| and |U| are fixed at 2000 and 4000, respectively. To better simulate the practical scenarios and to comprehensively evaluate all competing approaches, we use 3 different settings for γ = {0.2, 0.4, 0.6} to resemble the different proportions of positive samples within the unlabeled samples. ... for the CIFAR-10 benchmark, we adopt a convolutional neural network as the classifier; 2) for the remaining benchmarks, we adopt a multilayer perceptron (MLP) with a single-hidden-layer of 100 neurons as the classifier. |