전이학습을 이용한 시계열 데이터의 결측치 대체와 예측 성능과의 상관성 분석
A Transfer Learning for Missing Value Imputation and Its Relationship with Prediction Performance in Time Series Data
- 주제(키워드) Time Series Data , Missing Value Imputation , Transfer Learning , Prediction Accuracy , LSTM
- 주제(기타) 산업공학
- 설명문(URI) https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002985718
- 등재 KCI등재
- 발행기관 대한산업공학회
- 발행년도 2023
- URI http://www.dcollection.net/handler/ewha/000000238482
- 본문언어 한국어
- Published As http://dx.doi.org/10.7232/JKIIE.2023.49.4.294
초록/요약
Missing values incur the lack of data availability and/or inaccurate predictions in the problem of time series prediction. We consider a transfer learning method for missing data imputation in time series data and test two research hypothesis; the first hypothesis is that the high similarity between two time series, one containing missing values and the other used for transfer learning, improves the imputation performance. Second, a better imputation performance results in a better prediction accuracy. Empirical analysis reveals that the transfer learning with high similarity in two time series improves the imputation performance. As known in the literature, we found a positive correlation between imputation performance and prediction accuracy. However, the correlation between imputation performance and prediction accuracy becomes insignificant when the time series has low volatility and a short length of consecutive missing data. It means that a simple method for missing data imputation is preferred to an expensive but effective method such as transfer learning if the time series is highly stable and predictable.
more