Class Prior Estimation with Biased Positives and Unlabeled Examples

• DOI: 10.1609/aaai.v34i04.5848
• Publication Date: 2020-06-29T21:30:51Z
• Authors: Shantanu Jain; Justin Delano; Himanshu Sharma; Predrag Radivojac
• Keywords: Identifiability

Positive-unlabeled learning is often studied under the assumption that labeled positive sample drawn randomly from true distribution of positives. In many application domains, however, certain regions in support class-conditional are over-represented while others under-represented sample. Although this introduces problems all aspects positive-unlabeled learning, we begin to address challenge by focusing on estimation class priors, quantities central posterior probabilities and recovery classification performance. We start making a set assumptions model sampling bias. then extend identifiability theory priors unbiased biased setting. Finally, derive an algorithm for estimating relies clustering decompose original problem into subproblems learning. Our empirical investigation suggests feasibility correction strategy overall good