Feature Selection Method Based on Maximum Conditional and Joint Mutual Information

Abstract

In multidimensional time series data analysis, redundant features in samples can add to the complexity of problem analysis. Feature selection methods based on mutual information (MI) can effectively reduce data dimensionality and provide more accurate analysis results. Unfortunately, existing methods do not involve reasonable consideration to the relevance between time series features and rely on only one or two criteria to assess whether a feature is redundant. Given these problems, a time series feature selection method based on Maximum Conditional and Joint Mutual Information (MCJMI) is presented. It separates each time series into discrete ones. Two factors, overall Joint Mutual Information (JMI) and Conditional Mutual Information (CMI), are integrated in the selection of features. The experimental results demonstrate that MCJMI is both effective and useful for time series feature selection and can improve the accuracy and stability of feature selection.