Impute4TSC: Evaluating Missing Value Imputation Methods for Time Series Classification

Xiaoou Ding, Cong Peng, Muyun Zhou, Hongzhi Wang, Zhongyi Pei, Chen Wang, Jianmin Wang

Abstract: Time series classification is a crucial task in time series data analysis. Sensor-collected time series data have been observed to often contain missing values due to equipment failures, communication issues, and other factors that significantly affect the completeness of the data and the performance of the classification models. Although numerous missing value imputation methods for time series data have been proposed, our review reveals that existing methods inadequately consider the compatibility between imputation methods and time series classification models, thereby limiting their practical application effectiveness. This paper introduces Impute4TSC, an optimization framework for missing value imputation for time series classification. On one hand, we evaluate combinations of six imputation algorithms and seven classification algorithms using distribution consistency measures and classification performance metrics, and deeply analyze the impact of their compatibility on classification performance. On the other hand, based on the analysis of the interaction between imputation algorithms and classification models, we propose an optimized selection method for matching imputation and classification tasks, effectively improving the compatibility of missing value imputation methods.

External IDs:dblp:conf/icde/DingPZWPWW25