Date of Original Version
Abstract or Description
Active learning consists of principled on-line sampling over unlabeled data to optimize supervised learning rates as a function of the number of labels requested from an external oracle. A new sampling technique for active learning is developed based on two key principles: 1) Balanced sampling on both sides of the decision boundary is more effective than sampling one side disproportionately, and 2) exploiting the natural grouping (clustering) of unlabeled data establishes a more meaningful non-Euclidean distance function with respect to estimated category membership. Our new paired-sampling density-sensitive method embodying these principles yields significantly superior performance in multiple active learning data sets over all other sampling methods in our comparative study: representative sampling, uncertainty sampling, density-based sampling, and random sampling.