Comparison of Conventional and Machine Learning-Based (Modern) Imputation Methods in Estimating Lost Data in Time Series Data

Abstract

Time series data is very widely used in research. The obstacle that researchers often face is that it is difficult to produce accurate conclusions because there are missing data or imputation results are not good. Therefore, an appropriate imputation method is needed to address the lost data. In this study, the performance of spline interpolation and local average nearest neighbor (LANN) methods as conventional methods as well as recurrent neural networks (RNN) methods is modern. All three methods are simulated to address the different types of simulated data that are designed. The performance of these methods was evaluated with RMSE and MAE. The results obtained by the spline interpolation method work well on seasonal pattern data, then the LANN method can overcome trend and stationary data, while RNN works more flexibly and more accurately than the two conventional methods with RMSE 0.66 and MAE 0.53. So it can be concluded that the modern method of RNN works better than the interpolation of splines and LANNs. However, it should be noted that conventional methods can be superior in certain conditions.