PU-LP: A novel approach for positive and unlabeled learning by label propagation

Abstract

For the positive and unlabeled learning algorithms, when there is only small amount of labeled positive examples available, the algorithms can hardly extract reliable negative examples from the unlabeled examples in step one, which makes it hard to build the classifier with good performance in step two. Based on the same label assumption from graph based semi-supervised learning, we propose a novel graph-based PU learning algorithm, PU-LP, which takes Katz index to measure the similarities between vertices. After enlarging labeled positive set and extracting reliable negative examples, PU-LP build the classifier by label propagation algorithm. Experiments on UCI datasets shows that PU-LP has excellent performance when there is only small amount of labeled positive examples available, and it outperforms than PNB algorithm.