A Classification Method Using Transferring Shared Subspace

Transfer learning algorithms have been proved efficiently in pattern classification filed. The characteristic of transfer learning is to better use one domain information to improve the classification performance in different but related domains. In order to effectively solve the classification problems with a few labeled and abundant unlabeled data coming from different but related domains, a new algorithm named transferring shared subspace support vector machine (TS\+3VM) is proposed in this paper. Firstly a shared subspace used as the common knowledge between source domain and target domain is built and then classical support vector machine method is introduced to the subspace for the labeled data, therefore the resulting classification model has the ability of transfer learning. Specifically, using the theory of transfer learning and the principal of large margin classifier, the proposed algorithm constructs a shared subspace between two domains by maximizing the joint probability distribution of the labeled and unlabeled data. Meanwhile, in order to fully consider the distribution of the few labeled data, the classification model is trained in the augmented feature space consisting of the original space and the shared subspace. Experimental results confirm the efficiency of the proposed method.