Study Of Active Learning Algorithms Based On Uncertainty

With the widely application of computer and Internet, people can easily get abundant data. But for many practical applications, labeled instances are very difficult, expensive, or time-consuming to obtain, on the other hand unlabeled instances are easy to get. Active learning methods have been proposed to deal with such issues, it selects the most valuable unlabeled instances which then marked by the expert and added into training set. In this way, the active learner aims to achieve high accuracy using as few labeled instances as possible.A commonly used active learning framework is uncertainty sampling which selects instance with the least certain from current classifier. In this thesis, we analyze and discuss the methods of uncertainty based active learning which using nearest neighbor rules. This thesis presents weighted margin as the new uncertainty measure criterion for unlabeled instance. To avoid selecting isolated instance, we introduce the density measure for the most uncertain instance. The best instance should with high uncertainty and located at dense region. Finally, the experimental results on artificial and UCI datasets show that proposed selection strategy can achieve better performance than existing methods.