On Multiclass Active Learning with Support Vector Machines

In supervised machine learning, a training set of examples which are assigned to the correct target labels is a necessary prerequisite. However, in many applications, the task of assigning target labels cannot be conducted in an automatic manner, but involves human decisions and is therefore time-consuming and expensive. In the case of classification learning, the active learning framework has been considered to address this problem. While most research on active learning in the field of kernel machines has focused on binary problems, less attention has been given to the problem of learning classifiers in the case of multiple classes. We consider three common decomposition methods to express multiclass problems in terms of sets of binary classification problems and propose novel active learning heuristics in order to reduce the labeling effort. Various experiments conducted on real-world datasets demonstrate the merits of our approach in comparison to previous research.