Positive-Unlabeled Learning using Random Forests via Recursive Greedy Risk Minimization

Authors: Jonathan Wilton, Abigail Koay, Ryan Ko, Miao Xu, Nan Ye

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we compare PU Extra Trees with several other PU learning methods. Datasets We consider a selection of common datasets for classification from LIBSVM [5], as well as MNIST digits [19], the intrusion detection dataset UNSW-NB15 [25] and CIFAR-10 [17] to demonstrate the versatility of our method.
Researcher Affiliation Academia 1School of Mathematics and Physics, The University of Queensland 2School of Information Technology and Electrical Engineering, The University of Queensland 3RIKEN, Japan
Pseudocode Yes Algorithm 1: Learn DT(κ, S)
Open Source Code Yes Our code is available at https://github.com/puetpaper/PUExtra Trees.
Open Datasets Yes Datasets We consider a selection of common datasets for classification from LIBSVM [5], as well as MNIST digits [19], the intrusion detection dataset UNSW-NB15 [25] and CIFAR-10 [17] to demonstrate the versatility of our method.
Dataset Splits Yes : random 80%-20% train-test split was used as no train-test splits were provided.
Hardware Specification Yes In particular, random forests were trained using 32GB RAM and one of Intel i7-10700, Intel i7-11700 or AMD Epyc 7702p CPU. Neural networks were trained on one of NVIDIA RTX A4000 or NVIDIA RTX A6000 GPU due to the lack of identical devices.
Software Dependencies No The paper mentions general software categories like "neural networks" and references to "scikit-learn" implicitly through a citation, but it does not provide specific version numbers for any libraries, frameworks, or programming languages used (e.g., Python, PyTorch, TensorFlow, Scikit-learn versions).
Experiment Setup Yes Following common practice [26, 13], the default hyperparameters for PU ET are: 100 trees, no explicit restriction on the maximum tree depth, sample F = d features out of a total of d features and sample T = 1 threshold value when computing an optimal split. The architectures for the neural networks used in u PU, nn PU and Self-PU were copied from [16] for the 20News, epsilon, MNIST and CIFAR-10 datasets. A 6 layer MLP with Re LU was used for MNIST, Covtype, Mushroom and UNSW-NB15... For each dataset the neural networks were trained for 200 epochs. The batch size, learning rate, use of batch-norm..., weight decay and choice of optimiser were tuned for each dataset.