层次K均值聚类结合改进ITML的的迁移度量学习方法

目前的迁移学习方法多针对单一迁移类型，使用低级特征空间，并且源集比目标集复杂耗力；针对这些问题，综合考虑特征表示迁移、参数迁移和实例迁移，提出迁移度量学习的通用框架。首先，基于属性相似性空间和类别相似性空间，利用层次K均值聚类获取相似性；然后，利用去相关归一化转换方法消除源集中的相关关系来抑制负迁移作用；最后，改进信息理论度量学习方法进行相似性度量学习。对三种不同复杂度数据集进行实验，结果表明，提出方法的迁移学习性能较传统方法明显提高，且对负迁移影响具有更好的鲁棒性。此外，提出的方法可应用于源集比目标集简单的情况，评估结果表明，即使源集知识有限，也可以得到较好的迁移学习效果。

A Transfer Metric Learning Method Based on Hierarchical K-means Clustering and Improved ITML

Now most of transfer learning methods suffer from the problems that transfer types are separately analyzed, Low level feature space are used, and the source data set is more diverse and complex than the target set. For these problems, the paper proposes a novel general transfer metric learning framework with comprehensive consideration of feature representation transfer, parameter transfer and instance transfer. Initially, hierarchical k-means clustering is used to get the similarity based on the semantic similarity space and category similarity space. Then, the paper utilizes de-correlated normalized space to eliminate the correlation learned in the source domain, and restrain the negative transfer. Finally, we modify the information theoretic metric learning to precede similarity metric learning. The experiments have been done for three data sets with different complexity. The results show that the transfer learning performance of the proposed method has improved greatly with more robust to negative transfer effect comparing with the traditional methods. Furthermore, the proposed method could be applied in the situation that the source data set is simpler than the target set. The experiment results reveal that even when the knowledge source is limited, transfer learning can still be beneficial.