A BIC Based Initial Training Set Selection Algorithm for Active Learning and Its Application in Audio Detection

Abstract

To construct a classification system or a detection system, large amounts of labeled samples are needed. However, manual labeling is dull and time consuming, so researchers have proposed the active learning technology. The initial training set selection is the first step of an active learning process, but currently there have been few studies on it. Most active learning algorithms adopt random sampling or algorithms like sampling by clustering (SBC) to select the initial training samples. But these two kinds of method would lose their effectiveness in detecting events of small probability. Because sometimes they could not select or select too few samples of the small probability events. To solve this problem, this paper proposes a BIC based initial training set selection algorithm. The BIC based algorithm performs clustering on the whole training set first. Then uses BIC to judge the status of clusters. Finally, it adopts different selection strategies for clusters of different status. Experimental results on two real data sets show that, compared to random sampling and SBC, the proposed BIC based initial training set selection algorithm can efficiently solve the detection problem of small probability events. In the mean time, it has obvious advantages in detecting events of non-small probability.

Keywords

• Initial training set selection
• active learning
• BIC
• subspace sample selection
• audio detection