An active MBBNTree classifier learning from unlabeled samples

Abstract

Obtaining labeled training examples for some classification tasks is often expensive, such as text classification, mail filtering, while gathering large quantities of unlabeled examples is usually very cheap. Active learning aims at reducing the number of training examples to be labeled by automatically processing the unlabeled examples, then selecting the most informative ones with respect to a given cost function for a human to label. MBBNTree algorithm, which integrates the advantage of Markov Blanket Bayesian Networks (MBBN) and Decision Tree, would behave better performance than other Bayesian Networks for classification. But the available training samples with actual classes are not enough for building MBBNTree classifier in practice. In this paper, the MBBNTree classifier algorithm based on the Query-by-Committee of active learning would be presented to solve the problem of learning MBBNTree classifier from unlabeled samples. Experimental results show that the proposed algorithm can reach the same accuracy as passive learning with few labeled training examples.