A Hybrid Approach Based on Genetic Algorithm with Ranking Aggregation for Feature Selection

Abstract

Recently, feature selection has become challenging for many machine learning disciplines. The success of most existing approaches depends on the effectiveness of searching strategies to select the most salient features from the original feature space. Unfortunately, these approaches may become impractical when dealing with high-dimensional datasets. In order to overcome this problem, recent studies rely on a boundary scheme (i.e., fixing a number of selected features) to reduce the searching space or a ranking scheme (e.g., features with less correlated scores) to guide the selection phase. However, choosing the best-fitted size for the feature subset is also a hard problem, and relying on one feature comparison criteria may ignore important features. In this paper, we propose a genetic algorithm that aims to optimize the feature subset and the appropriate number of selected features to maximize the performance of an Artificial Neural Network (ANN) classifier. To improve the efficiency of the selection phase, we combine the proposed GA with a local search algorithm based on a ranking aggregation approach. Our objective is to speed up the searching algorithm by taking advantage of different feature scoring criteria. We have assessed the performance of our approach over three categories of datasets: small, medium and high in terms of feature dimensionality (e.g., the smallest and the largest datasets include 8 and 7129 features, respectively). The empirical results have shown that our proposed approach outperforms the other state-of-the-art works when dealing with medium- and high-dimensional datasets and is comparable to them in the case of small-dimensional datasets.