Exploiting Relations Among Concepts to Acquire Weakly Labeled Training Data

Abstract

We consider a learning setting in which there are well de ned relations that exist among instances of certain classes. In particular, we consider the domain of predicting various types of gene-regulation elements in bacterial genomes. Given instances of one class, we can often acquire \\weakly labeled" training data for another class by taking advantage of known relationships that exist between the two classes. The examples are weakly labeled in that either the class label is incompletely speci ed, the exact extent of an instance is only partially speci ed, or both. We use an EM-based approach to handle this hidden state during learning. Our experimental results show that, when only small training sets are available, there can be signi cant value in augmenting the training sets with weakly labeled examples acquired from relationships among concepts.