A new domain adaption algorithm based on weights adaption from the source domain

作者：Lu, Nannan^*; Chu, Fei; Qi, Haoran; Xia, Shuang

来源：IEEJ Transactions on Electrical and Electronic Engineering, 2018, 13(12): 1769-1776.

DOI：10.1002/tee.22739

• domain adaption
• feature representation
• instance transfer
• weights adaption
• distribution adaption

Domain adaption is effective for utilizing the knowledge from a source domain to benefit further learning task in a different but related target domain. It mainly reduces the difference across domains by learning a shared feature representation. So, it also could be taken as a feature transfer algorithm. Besides, another transfer technique, named instance transfer, allows the instances from a source domain to be reused in a target domain by reweighting. Each method has two facets. Instance transfer must satisfy the high similarity between the source domain data and the target domain data, while domain adaption ignores the difference of contributions by the different instances. In this paper, a new domain adaption algorithm is proposed to introduce weights adaption into the feature representation. Specifically, it aims to reweight the instances in new feature space and adapt the distributions across the different domains. In distribution adaption, both the marginal distribution and the conditional distribution are simultaneously adjusted in the unified framework. In this way, the new feature representation is effective for the substantial distribution difference as well as instance difference. The experiments on 36 groups of image datasets show that the knowledge of the source domain is efficiently transferred to classify the target domain.

• 出版日期2018-12
• 单位中国矿业大学（北京）