Two-Stage Angular Alignment for Positive-Unlabeled Learning

Positive-Unlabeled (PU) learning addresses the binary classification problem where only positive and unlabeled data are available—a setting common in applications such as medical diagnosis and web mining. We introduce a novel two-stage approach based on angular alignment in feature space, where a learnable prototype vector represents the directional centroid of the positive class. In the first stage, the model aligns labeled positives toward this prototype to promote angular compactness; in the second, it repels overly similar unlabeled instances to refine the decision boundary without prematurely assigning negative labels. Our method employs a directional loss inspired by von Mises–Fisher geometry, a dynamic stage-switching curriculum, and maintains a highly parameter-efficient design. Experiments on CIFAR-10 and SVHN demonstrate strong performance and competitive results compared to state-of-the-art PU learning methods. The approach also yields semantically structured latent spaces, highlighting the value of angular geometry for interpretable and effective representation-based PU learning in visual domains.