A novel hybrid feature selection based on ReliefF and binary dragonfly for high dimensional datasets

Abstract

High dimensionality is a common challenge in large datasets. Combination of the filter and wrapper methods is used to select the appropriate set of features in these datasets. The hybrid method is desirable, which uses the advantages of both the methods and covers the disadvantages. In this paper, a hybrid method for feature selection in high dimension data is presented. In proposed algorithm, the ReliefF algorithm is used as a filter method for ranking features. Next, the binary dragonfly algorithm (BDA) is applied as a wrapper method. The BDA algorithm uses the ranked features to find optimal set of features incrementally and iteratively. Minimizing the cross-validation loss and decreasing the number of features is considered to evaluate the solution, hierarchically. The proposed algorithm and other compared algorithms run over 5 datasets and the results indicated that the proposed algorithm not only reduce the dimension of dataset but also improve the performance of classifiers on the test data.