부분공간 정렬을 통한 비지도 도메인 적응 분류 모형 비교 분석

Abstract

Unsupervised domain adaptation (DA) aims to build a classification model that performs well in the target domain by utilizing some data or information from a source domain that is different but associated with the target domain and has sufficient labeled instances. In this paper, a comparative study on unsupervised domain adaptation methods using subspace alignment was conducted. Two main algorithms are considered. One is the subspace alignment DA and the other is the subspace distribution alignment DA. They are based on dimension reduction using PCA (principal component analysis) and basis vector alignment of the subspace or distribution alignment of the subspace. In addition, a method for realizing domain adaptation through subspace alignment while utilizing supervised PCA for subspace derivation was proposed and compared with the previous two methods. In an image classification data analysis, it was confirmed that the subspace distribution alignment domain adaptation method showed superior results than the subspace alignment domain adaptation, and the proposed domain adaptation method utilizing a supervised PCA and subspace alignment showed the best performance in some situations.