Reviews: Positive-Unlabeled Learning with Non-Negative Risk Estimator

Summary The paper builds on the literature of Positive and Unlabeled (PU) learning and formulates a new risk objective that is negatively bounded. The need is mainly motivated by the fact that the state of the art risk formulation for PU can be unbounded from below. This is a serious issue when dealing with models with high capacity (such as deep neural networks), because the model can overfit. A new lower-bounded formulation is provided. Despite not being unbiased, the new risk is consistent and its bias decreases exponentially with the sample size. Generalization bounds are also proven.