Research On Change-Rate-based Stop Criteria Of Active Learning Algorithms

Active learning, which becomes a hot topic in machine learning research area, aims to train a learner with high performance on the condition that we just need to label instances as less as possible whhen there are few samples with class labels. It has been whidely applied to many application areas such as image retrival, documents classification and DNA identification. Many researchers devote themselves to the improvement of the sample selection strategies for active learning algorithms and have made great achievements. But the research on the stop criteria of active learning is unseen, which is the research topic of the thesis.Since the performance of the learner does not become higher all the time as the more and more samples are added in training set, thus how to balance the two aspects becomes a troublesome issue. The thesis introduces the change rate of the prediction labels in the pool to evaluate the performance of the current learner. When the change rate is very large, it means that the learner is not enough and it needs to learn continoue; when the change rate is very small, it is means that the learner is good and robust enough. When the change rate is smaller than a threshold given beforehand, the sample selection process can be stopped. The numerical experimental results on UCI show that the new stop criterion can get a learner with higher performance while less cost for annotation is needed than existing stop criteria.