基于域间相似度序数的迁移学习源领域的选择

现有迁移学习研究大多数都建立在源领域和目标领域的相似度较高的全局约束下,对如何选择合适的源领域缺乏研究。为了确定如何自适应地从候选源领域集合中选择合适源领域,提升迁移效率,避免"负迁移"现象,基于最大均值差异(maximum mean discrepancy,MMD)提出一种叫作域间相似度序数(MMD-SR)的度量方法,用于度量候选源领域与目标领域间的相似度。同时,基于MMD-SR,提出一种迁移学习源域自适应选择策略(MMD-SR source domain selection strategy,MMD-SR_SDSS)。在人工数据集和真实数据集中的实验结果表明了度量方法MMD-SR和源领域选择策略MMD-SR_SDSS的有效性和可行性。

Source Domain Selection in Transfer Learning Based on Domain Similarity Rank

Most of the existing transfer learning studies are based on the global constraint that the similarity between source domain and target domain is high. There is a lack of research on how to choose the appropriate source domain. In order to determine how to select the appropriate source domain from the candidate source domain set adaptively, improve the transfer efficiency and avoid the phenomenon of "negative transfer", this paper proposes a measurement method called domain similarity rank(MMD-SR) based on maximum mean discrepancy(MMD) to measure the similarity between the candidate source domain and the target domain. At the same time, an adaptive source domain selection strategy MMD-SR_SDSS is proposed which based on MMD-SR. The effectiveness and feasibility of MMD-SR and MMD-SR_SDSS are illustrated by experiments on artificial datasets and real datasets.