Machine Learning for Univariate Time Series Imputation
Thi-Thu-Hong Phan
Abstract
Missing data occur in almost real time series applications. Using incomplete data or ignoring missing values can cause inaccurate results and reduce system efficiency. Recovering missing data so plays an important role in time series analysis but this task still remains a challenge. Therefore, this study aims to propose a new approach for filling consecutive missing values (gap) in univariate time series using machine leaning (ML) methods, namely MLBUI. Firstly, for each gap we transform the data before the gap and the data after this gap into multivariate time series. After this transformation, forward and backward forecasting based on ML methods are applied to estimate missing values. Finally, we impute the gap by average values of the both forecast sets. Four real-world datasets are used for assessing the performance of the proposed approach in comparison with five other imputation methods using five quantitative and visualization metrics. Experimental results show that MLBUI approach outperforms than several state-of-the art methods.