WBDI Approach for Univariate Time Series Imputation

Abstract

Incomplete data can significantly impact the results and reduce data value for machine learning systems. Simple imputation methods often fail to capture the intricate patterns and relationships within time series data, leading to inaccurate analysis. This study proposes a novel approach called ”weighted bi-directional imputation, WBDI” to address this challenge in univariate time series data. The proposed approach leverages machine learning models and utilizes data from both before and after the missing segment, incorporating weights to prioritize relevant information. To evaluate its effectiveness, experiments are conducted using eleven machine learning algorithms on three real-world datasets with varying sizes and sampling frequencies. The results demonstrate that ensemble learning methods generally outperform other approaches. Notably, the AdaBoost method consistently achieves top performance across all datasets and evaluation metrics,
illustrating its high reliability and accuracy in imputing missing values.