Angular Regularization for Positive-Unlabeled Learning on the Hypersphere
Authors: Vasileios Sevetlidis, George Pavlidis, Antonios Gasteratos
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on benchmark datasets demonstrate that Angular PU achieves competitive or superior performance compared to state-of-the-art PU methods, particularly in settings with scarce positives and high-dimensional embeddings, while offering geometric interpretability and scalability. |
| Researcher Affiliation | Academia | Vasileios Sevetlidis* EMAIL Athena RC Institute for Language and Signal Processing University Campus Kimmeria Xanthi, GR67100, Greece Democritus University of Thrace Dept. Production and Management Engineering Vas. Sofias 17 Xanthi, GR67100, Greece * Corresponding Author George Pavlidis EMAIL Athena RC Institute for Language and Signal Processing University Campus Kimmeria Xanthi, GR67100, Greece Antonios Gasteratos EMAIL Democritus University of Thrace Dept. Production and Management Engineering Vas. Sofias 17 Xanthi, GR67100, Greece |
| Pseudocode | No | The paper describes the proposed method in Section 3 'Proposed Method' using mathematical equations and descriptive text, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate the proposed v MF PU learning framework on a range of benchmark datasets: CIFAR-10, STL10, SVHN, and ADNI. These datasets span natural images, digit recognition, and medical imaging, allowing us to assess performance across varying input complexity and domain characteristics. ...We evaluate the proposed Angular PU method in a PU regime on two large Med MNIST datasets. |
| Dataset Splits | Yes | To ensure comparability with prior work, we replicate the experimental setup of Yuan et al. (2025), including dataset partitions, positive/unlabeled (PU) label ratios, and evaluation metrics. ...Unless otherwise stated, all experiments use a validation split (10% of the training data) to tune the decision threshold for binary classification by maximizing F1 score. ...CIFAR-10: We adopt the standard 50,000/10,000 train/test split. ...STL-10: A total of 500 positive examples are randomly selected as labeled data, with the rest used as unlabeled. ...SVHN: We randomly sample 1,000 even digits as labeled positives, while treating the remainder of the training set as unlabeled. The test set includes 26,032 labeled images. ...ADNI: From the positive class, 768 scans are randomly selected as labeled data. ...Med MNIST: From the pooled positives, we randomly sample up 12,5% labeled positives (cap applied per seed); all remaining samples both positives and negatives are treated as unlabeled. The official test split is used only for evaluation. Across S = 10 seeds (Table 6), we re-draw the labeled-positive subset and re-train to quantify variability due to the PU sampling. |
| Hardware Specification | No | The paper does not provide specific hardware details (like GPU or CPU models) used for running the experiments. It only describes the models, optimizers, and general training parameters. |
| Software Dependencies | No | The paper mentions using the Adam optimizer but does not specify its version or any other software dependencies (e.g., libraries, frameworks) with version numbers. |
| Experiment Setup | Yes | For image datasets (CIFAR-10, STL-10, SVHN), we adopt a VGG11-BN encoder (Simonyan & Zisserman, 2014) pretrained on Image Net (Deng et al., 2009). For the ADNI dataset (Jack Jr et al., 2008; Petersen et al., 2010), we use a two-layer multilayer perceptron (MLP) with Re LU activations. All models are trained for 15 epochs using the Adam optimizer with a learning rate of 10 4 and a batch size of 128. The encoder maps inputs to a d-dimensional hypersphere, with d = 128 unless stated otherwise. We apply dropout with a rate of 0.2 to the final embedding layer. The cosine uniformity regularizer is scaled by temperature t = 2.0. The concentration parameter κ of the von Mises-Fisher distribution is tuned per dataset as a hyperparameter. Unless otherwise specified, all components of the encoder are updated end-to-end, including learnable margins where applicable. |