DEDPUL: Method for Positive-Unlabeled Learning based on Density Estimation

Positive-Unlabeled Classification is an analog of binary classification for the case when the Negative (N) sample in the training set is contaminated with latent instances of the Positive (P) class and hence is Unlabeled (U). We develop DEDPUL, a novel method that simultaneously solves two problems concerning U: estimates the proportions of the mixing components (P and N) in U and classifies U. We conduct experiments on synthetic and real-world data and show that DEDPUL outperforms current state-of-the-art methods for both problems. We suggest an automatic procedure for the objective choice of DEDPUL hyperparameters. Additionally, we improve two methods from the literature.