Selective Supervision: Guiding Supervised Learning with Decision-Theoretic Active Learning

Ashish Kapoor, Eric Horvitz, Sumit Basu

An inescapable bottleneck with learning from large data sets is the high cost of labeling training data. Unsupervised learning methods have promised to lower the cost of tagging by leveraging notions of similarity among data points to assign tags. However, unsupervised and semi-supervised learning techniques often provide poor results due to errors in estimation. We look at methods that guide the allocation of human effort for labeling data so as to get the greatest boosts in discriminatory power with increasing amounts of work. We focus on the application of value of information to Gaussian Process classifiers and explore the effectiveness of the method on the task of classifying voice messages.