An Online Transfer Learning Algorithm with Adaptive Cost

Online transfer learning aims to attack an online learning task on a target domain by transferring knowledge from some source domains, which has received more attentions. And most online transfer learning methods adapt the classifier according to its accuracy on new coming data. However, in real-world applications, such as anomaly detection and credit card fraud detection, the cost may be more important than the accuracy. Moreover, the cost usually changes in these online data, which challenges state-of-art-methods. Therefore, this paper introduces the cost of misclassification into transfer-learning of classifier, and proposes a novel online transfer learning algorithm with adaptive cost (OLAC). Firstly, we introduce the label distribution into traditional Hinge Loss Function to compute the cost of classification adaptively. Secondly, we transfer learn the classifier according to its performance on new coming data including both accuracy and cost. Extensive experimental results show that our method can achieve higher accuracy and less classification lost, especially for the samples with higher costs.