Algorithmic cache of sorted tables for feature selection: Speeding up methods based on consistency and information theory measures

Abstract

Feature selection is a mechanism used in Machine Learning to re-duce the complexity and improve the speed of the learning process by usinga subset of features from the data set. There are several measures which areused to assign a score to a subset of features and, therefore, are able to com-pare them and decide which one is the best. The bottle neck of consistencemeasures is having the information of the different examples available to checktheir class by groups. To handle it, this paper proposes the concept of an al-gorithmic cache, which stores sorted tables to speed up the access to exampleinformation. The work carries out an empirical study using 34 real-world datasets and four representative search strategies combined with different tablecaching strategies and three sorting methods. The experiments calculate fourdifferent consistency and one information measures, showing that the proposedsorted tables cache reduces computation time and it is competitive with hashtable structures.