Transductive Transfer Classification With Active Learning From Source Domain

While there are a small amount of labeled samples,saving labeling efforts is an important research topic in the area of machine learning community.In order to obtain an efficient classifier,we can adopt transfer classification to transfer knowledge from a related auxiliary domain,or active learning to query the most informative samples for labeling.However,in many real applications,existing methods pay more attention to the situation that labeled samples and the oracle is available in the target domain.In this paper,we consider a more challenging situation that labeled samples and the oracle are totally unavailable in the target domain.We combine active learning and transductive transfer classification,and propose an efficient and novel algorithm,called transductive transfer classification with active learning from source domain(TTCALS),to utilize the knowledge from source domain.Our key idea is to class-wisely select the most informative samples in the source domain and learn the shared subspace to transfer the knowledge from source domain,both these steps are integrated into one united framework.Considering the representativeness and information in the active query,we combined it with the regularization term in the objective function,and proposed informative transdutive transfer classification with active learning from source domain(RITTAL)method.The main contributions of this work is as follows:? We study a challenging problem setting where labeled samples is only present andinfrequent in the source domain and no oracle is available in the target domain.Byselecting and labeling a small amount of sample from source domain,which aremost representative and informative,it is helpful to predict the sample from targetdomain.? To the best knowledge,this is the first work to combine active learning and trans-ductive transfer learning.Samples in source domain are actively selected and as-signed label by experts.And the shared subspace between source domain andtarget domain is learned by transfer learning algorithm.Based on this method,theknowledge between domains can be basically transferred.? Besides,we propose our basic method that the name is transductive transfer classi-fication with active learning from source domain(TTCALS).considering the regu-larization of objective function and representativeness in active query,we revise ourmethod TTCALS,propose the representative and informative transdutive transferclassification with active learning from source domain(RITTAL)method.? Experiments on the Reuters-21578 and 20-Newsgroup demonstrate that the pro-posed methods,TTCALS and RITTAL,outperform the state-of-the-art related meth-ods,and can save labours at the same time.