Active Learning with Partially Labeled Data via Bias Reduction

With active learning the learner participates in the process of selecting instances so as to speed-up convergence to the “best ” model. This paper presents a principled method of instance selection based on the recent bias variance decomposition work for a 0-1 loss function. We focus on bias reduction to reduce 0-1 loss by using an approximation to the optimal Bayes classifier to calculate the bias for an instance. We have applied the proposed method to naïve Bayes learning on a number of bench mark data sets showing that using this active learning approach decreases the generalization error at a faster rate than randomly adding instances and converges to the optimal Bayes classifier error obtained from the original data set.