Evolutionary Instance Selection Based on Preservation of the Data Probability Density Function

Abstract

The generation of massive amounts of data has motivated the use of machine learning models to perform predictive analysis. However, the computational complexity of these algorithms depends mainly on the number of training samples. Thus, training predictive models with high generalization performance within a reasonable computing time is a challenging problem. Instance selection (IS) can be applied to remove unnecessary points based on a specific criterion to reduce the training time of predictive models. This paper introduces an evolutionary IS algorithm that employs a novel fitness function to maximize the similarity of the probability density function (PDF) between the original dataset and the selected subset, and to minimize the number of samples chosen. This method is compared against six other IS algorithms using four performance measures relating to the accuracy, reduction rate, PDF preservation, and efficiency (which combines the first three indices using a geometric mean). Experiments with 40 datasets show that the proposed approach outperforms its counterparts. The selected instances are also used to train seven classifiers, in order to evaluate the generalization and reusability of this approach. Finally, the accuracy results show that the proposed approach is competitive with other methods and that the selected instances have adequate capabilities for reuse in different classifiers.