Exploiting Structural Information Of Data In Active Learning

In recent years, the active learning algorithms have focused on combining correlation criterion and uncertainty criterion for evaluating instances. Although these criteria might be useful, applying these measures on whole input space globally may lead to inefficient selected instances for active learning. The proposed method takes advantage of clustering to partition input space to subspaces. Then it exploits both labeled and unlabeled data locally for selection of instances by using a graph-based active learning. We define a novel utility score for selecting clusters by combining uncertainty criterion, local entropy of clusters and the factor of contribution of each cluster in queries. Experimental results reveal an elevated performance as compared to several state of the art and widely used active learning strategies.