An iterative algorithm to learn from positive and unlabeled examples

Xin Liu, Qingle Zheng, Xiaotong Shen, Shaoli Wang
2022 Statistica sinica  
In semi-supervised learning, a training sample is comprised of both labeled and unla-2 beled instances from each class under consideration. In practice, an important yet challenging 3 issue is the detection of novel classes that may be absent from the training sample. In this article, 4 we focus on a binary situation in which labeled instances come from the positive class whereas 5 unlabeled instances from both classes. Particularly, we propose a semi-supervised large margin 6 classifier to
more » ... 6 classifier to learn the negative (novel) class based on pseudo-data iteratively generated using 7 an estimated model. Numerically, we employ an efficient algorithm to implement the proposed 8 method with the hinge-loss and ψ-loss functions. Theoretically, we derive a learning theory for 9 the new classifier to quantify the misclassification error. Finally, numerical analysis demonstrates 10 that the proposed method compares favorably on simulated examples and is highly competitive 11 against its competitors on benchmark examples. 12
doi:10.5705/ss.202020.0287 fatcat:uab2nvzgabavrgn5k3mtlgvnsi