Learning with hybrid data

Learning with hybrid data aims at inducing a classifier that learns from partly labeled data. In this paper, four semi-supervised learning (SSL) methods are discussed. These include clustering with partial supervision, active sampling for learning with RBF networks, Gaussian mixture models based on the EM method, and finally seed-based clustering. The empirical study shows that the effect of unlabeled data on the accuracy for some algorithms is significant, while that of others depends on the data and the assumptions underlying the algorithms themselves.