Computing Mutual Information of Big Categorical Data and Its Application to Feature Grouping

Abstract

This paper develops a parallel computing system - MiCS - for mutual information of big categorical data on the Spark computing platform. The MiCS algorithm is conductive to processing a large amount and strong repeatability of mutual-information calculation among feature pairs by applying a column-wise transformation scheme. And to improve the efficiency of the MiCS and the utilization rate of Spark cluster resources, we adopt a virtual partitioning scheme to achieve balanced load while mitigating the data skewness problem in the Spark Shuffle process.