Feature Selection through Filtering with Mono and Multi-Objective Memetic Algorithms Using Correlation

Feature selection is the process of extracting the most relevant features from a dataset, helping to reduce its dimensionality by eliminating non-essential features. This leads to simpler, faster models and optimises training efficiency. This paper presents two memetic algorithms: one employs a mono-objective filter method as a fitness function, while the other adopts a multi-objective approach. The latter uses the number of attributes in the dataset as the first objective, and the sum of Pearson’s correlations for the selected attributes as the second. Additionally, we apply a novel approach to the use of correlation for attribute selection within the aforementioned memetic algorithms. Both proposals aim to identify the most relevant attributes to reduce the dimensionality of twelve test datasets. The performance of the selected features was evaluated using a J48 decision tree. The results showed a reduction in the number of attributes ranging from 14% down to 5%, while accuracy varied from −5% up to 11%, with an average improvement of over 4% (considering only those datasets where accuracy changed).