Reviews: Binary Classification from Positive-Confidence Data

Overview and Recommendation: This paper presents an algorithm and theoretical analysis for Pconf learning of a binary classifier, when only the positive instances and their conditional class probabilities are available. In contrast to PU learning, which has received attention recently in the ML community, Pconf learning does not require a large amount of unlabeled data or any knowledge of class priors, however it does require confidence weights p(y 1 x) for each example x sampled from p(x y 1). The paper is well written and this solution can be useful for many real world applications, therefore, it is a clear accept. Although, the analysis is very similar to the recent analysis for PU learning published by Du Plessis et al. recently at NIPS but the problem setting is novel and interesting. However, I think that the authors should fix some technical issues with their simulation experiment.