Abstract

In this paper, we consider active learning as a procedure of iteratively performing two steps: first, we train a classifier based on labeled and unlabeled data. Second, we query labels of some data points. The first part is achieved mainly by standard classifiers such as SVM and logistic regression. We develop additional techniques when there are very few labeled data. These techniques help to obtain good classifiers in the early stage of the active learning procedure. In the second part, based on SVM or logistic regression decision values, we propose a framework to flexibly select points for query. We find that selecting points with various distances to the decision boundary is important, but including more points close to the decision boundary further improves the performance. Our experiments are conducted on the data sets of Causality Active Learning Challenge. With measurements of Area Under Curve (AUC) and Area under the Learning Curve (ALC), we find suitable methods for different data sets.