Online transfer learning with multiple source domains for multi-class classification

Abstract

The major objective of transfer learning is to handle the learning tasks on a target domain by utilizing the knowledge extracted from the source domain(s), when the labeled data in the target domain are not sufficient. Transfer learning can be classified into offline transfer learning (OffTL) and online transfer learning (OnTL), and OnTL has attracted much attention and research due to its more realistic scenario assumed in practice. There can be multiple source domains, therefore, OnTL with Multiple Source Domains has been studied in recent years and algorithms have been proposed. Nevertheless, it can be noted that existing research on OnTL with Multiple Source Domains only deals with binary classification tasks. In this paper, we make the first attempt to study OnTL with Multiple Source Domains for multi-class classification (MC), and propose an algorithm, referred to as Online Multi-source Transfer Learning for Multi-class classification (OMTL-MC) algorithm. OMTL-MC algorithm is built on two-stage ensemble strategy, in this way, the knowledge extracted from different source domains can be simultaneously online transferred to improve the performance of the classifier in the target domain. In order to deeper explore the underlying structure among multiple classes, an Extended Hinge Loss (EHL) function is adopted in OMTL-MC. We theoretically analyze the mistake bound of OMTL-MC algorithm. In addition, experiments on several popular datasets expound that the proposed OMTL-MC algorithm outperforms the other compared algorithms.