An Ensemble Approach to Domain Adaptation in Sentiment Analysis

• 발행기관: 한국자료분석학회
• 저자명: 연규필(Kyupil Yeon)
• 간행물 정보: 『Journal of The Korean Data Analysis Society (JKDAS)』Vol.21 No.4, 1645~1653쪽, 전체 9쪽
• 주제분류: 자연과학 > 통계학
• 파일형태: PDF
• 발행일자: 2019.08.31

영문 초록

Domain adaptation aims to predict a response variable accurately in a target domain where labeled instances are non-exist or scarce, if any, but unlabeled instances are plentiful, by utilizing as much as possible abundant labeled information in one or several related source domains. Therefore it can be considered as a transfer learning which is one of hot topics in machine learning community. Although a target domain is related to source domains in any way, the underlying distributions generating instances are different. This inevitably carries on the so called concept drift phenomenon which means the input-output dependency changes across the domains. Under the concept drift, many ensemble learning algorithms have been suggested and showed quite good results. In this paper, we apply an ensemble learning scheme for semi-supervised domain adaptation as in the concept drift learning where a penalized regression based ensemble combiner is utilized. The proposed method is applied to a sentiment classification and shows a good result.

목차

1. Introduction

2. Ensemble based concept drift learning

3. Ensemble based domain adaptation algorithm

4. Data Analysis

5. Conclusions

References

키워드

domain adaptation ensemble learning transfer learning concept drift