Researches On Cross-domain Transfer Learning

Data mining and machine learning technologies have already achievedsignificant success in many knowledge engineering areas. However, many machinelearning methods work well only under a common assumption: the training and testingset are draw from the same domain and the same distribution. But in general, thetraining data and testing data are from different domains, thus, traditional machinelearning technology will not be able to obtain the very good learning effect. If we canuse a domain’s knowledge to help the study of another domain, namely thecross-domain transfer learning, then the effect of machine learning algorithm will begreatly improved.This thesis studies the classification of transfer learning and introduces sometypical transfer learning algorithm. Based on these, this thesis proposes two newtransfer learning algorithms:（1）This thesis proposes a parameter-transfer learning method based on KernelMean Matching algorithm. First reweight each source instance using KMM and thenapply the reweighted instances to the learning method based on parameters.Experimental results demonstrate that the proposed method outperforms the methodswhich based on instances or the parameters, especially when the number of targettraining data is relatively few.（2）This thesis proposes a semi-supervised learning transfer algorithm based onKDA. The KDA maps the marginal distribution of target domain and source domaindata into a common kernel space. In this common feature space, the differencebetween the two domains can be reduced. Then utilize the improved Co-trainingmethod to update the weights of the source domain’s data and get two learners for thelearning of target domain. The experimental evaluations were performed on theReuters21578and20Newsgroup, which are two well-known data sets in the field ofdocument classification. Experimental results show that the proposed algorithm canachieves higher classification accuracy.