中文摘要：

针对在单一匹配边缘概率分布以缩减源域和目标域的差异性时存在的泛化能力差的问题，提出联合边缘概率分布和条件概率分布减小域间差异性的基于特征和实例的迁移学习算法,通过核主成分分析在子空间中寻找样本新的特征表示，在该子空间中利用最小化最大均值差异，联合匹配边缘概率分布和条件概率分布以减小源域和目标域间的差异性．同时利用L2,1范数约束选择源域中相关实例进行训练，进一步提高迁移学习获得的模型泛化性能．在字符集和对象识别数据集上的实验表明文中算法的有效性．

英文摘要：

Aiming at the poor generalization ability of only matching marginal probability distribution to reduce the domain difference, a feature joint probability distribution and instance based transfer learning algorithm （FJPD-ITLA） is proposed. The instances are represented with the kernel principal component analysis in subspace. In this subspace, the maximum mean discrepancy is expanded to jointly match the marginal and conditional probability distribution. Thus, the difference between the source domain and target domain is reduced. Meanwhile, the L2,1-norm constraint is utilized to choose relevant instances in the source domain, and the generalization ability of the model obtained by transfer learning is improvedfurther. Experimental results on the digital and object recognition datasets demonstrate the validity and efficiency of the proposed algorithm.