Active learning with partially-labeled data to reduce classification loss.

摘要

In many occasions in real life, we are faced with the problem of classification of partially labeled data, or semi-supervised learning. We consider the special case of scarcely labeled data or when the labeled data is insufficient, and present a principled method which implements active learning in scarcely labeled data to enhance the performance of the learner. Our active learning algorithm is a co-training procedure that does not require natural view splits. Blum and Mitchell's co-training is a popular semi-supervised algorithm to use when we have multiple independent views of the entities to classify. An example of a multi-view situation is classifying web pages: one view may describe the pages by the words that occur on them, another view describes the pages by the words in the hyperlinks that point to them. So we start by analyzing co-training procedure and then introduce our general active learning algorithm.; This method is based on the recent bias variance decomposition work for a 0-1 loss function. We focus on bias and variance reduction to reduce 0-1 loss by first selecting a random pool from the unlabeled data, and then using the most-informative instances from that pool to reduce the variance, bias, and thereby overall loss of the learner via active learning.; Finally we focus on the best instance selection for labeling the unlabeled data; we use Jensen-Shannon divergence as one selection criterion. We show that our single instance selection approaches are superior to multiple instance selection approach. Our empirical results show that this technique can improve the generalization error with less running time compared with other active learning algorithms.