关键词: 多源迁移学习, 在线学习, 不平衡数据, 特征空间, 支持向量机（SVM）, k近邻, 核函数

Abstract: Multi-source online transfer learning uses the labeled data of multiple source domains to enhance the classification performance of the target domain. Aiming at the imbalanced datasets, this paper proposes a multi-source online transfer learning algorithm which can oversample in the feature space of the source domain and the target domain. The algorithm consists of two parts: oversampling multiple source domains and oversampling online target domains. In the oversampling stage of source domain, the feature space of the support vector machine (SVM) classifier is oversampled to generate a few class samples. The new samples are obtained by amplifying the original Gram matrix from the neighborhood information in the feature space of the source domain. In the oversampling phase of target domain, the samples of target domain arrive in batches, and a few samples of current batch look for the k-nearest neighbor in the feature space from the previous batches, and the target domain function is trained with the generated new samples and the original samples of current batch. Through the kernel function, the samples of the source domain and the target domain are mapped into the same feature space for oversampling, and the corres-ponding decision functions are trained by using the data of source domain and target domain with relatively balan-ced class distribution, so as to improve the overall performance of the algorithm. Experiments are carried out on four real datasets. On the widely used Office-Home dataset, the accuracy is improved by 0.0311 and the G-mean value is improved by 0.0702 compared with other baseline algorithms.

Key words: multi-source transfer learning, online learning, imbalanced data, feature space, support vector machine (SVM), k-nearest neighbor, kernel function