Fast Cross-Domain Classification Method for Large Multisources/Small Target Domains

Most of current cross-domain classifiers are proposed for single source and single target domains and basically based on the assumption that there is a balance between these two domains. However, this assumption is often violated in the real world. When these classifiers are applied to imbalanced domains, their classification performance and robustness to noise will heavily degrade. For example, Baysian classifier depends heavily on the estimation of the sample distributions of source and target domains. When large source domain but only a small target domain are available, the classification accuracy of this classifier will degrade a lot. In order to address this imbalanced issue and use abundant data in the source domain to do an effective transfer learning between small target domain and multisource domains, a novel fast cross-domain classification method called IMCCL for “small-target+multisource” datasets is proposed here. The proposed method IMCCL is rooted at logistic regression model and MAP. Accordingly, the proposed IMCCL is integrated together with the latest advance—CDdual algorithm—to develop its fast version IMCCL-CDual for “small-target+large-multisource” domains. This fast classification method is also theoretically analyzed. Our experimental results on artificial and real datasets indicate the effectiveness of the proposed method IMCCL-CDual in classification accuracy, the classification speed, robustness and domain adaption.