Active Learning Using a Low-Rank Classifier

Abstract

The majority of learning algorithms work based on a training dataset. However, labeling the collected data is costly and time consuming. Active learning has gained high attention due to its ability to label a vast amount of unlabeled collected data. However, the performance of the current state-of-the-art methods declines when the number of training data is increasing. In this paper, we propose and study a variant of Support Vector Machine (SVM), namely low-rank classifier, which is regularized by the trace-norm of learning parameters in active learning scenario. We compare this algorithm with the standard SVM algorithms in depth and analyze its computational complexity and optimization solution. Our experimental results confirm, that the proposed method outperforms the other methods for an increasing amount of training data.