An improved parallel matrix factorization method for data imputation of multivariable time series data with high level noises

摘要

Large number of multivariable time series data with high level noises are generated during the industrial production process. However, due to various uncontrollable reasons in the production process, the data missing problem often occurs. Considering that there may be potential correlation between the variables of multivariable time series, and the single variable will not fluctuate dramatically over time, an improved matrix factorization method was proposed in this paper to impute the missing data of multivariable time series. Due to the efficiency requirement of data imputing in industrial production, a dynamic updating method of parameters was proposed. In addition, filter matrix and improving correlation judgment was proposed to increase the imputing accuracy. Experiments on real datasets and self-made datasets show that the proposed method can effectively improve the imputing accuracy and efficiency.