Multi-objective optimization of feature selection using hybrid cat swarm optimization

With the pervasive generation of information from a wide range of sensors and devices, there always exist a large number of input features in databases, thus complicating machine learning problem formulation. However, certain features are relatively impertinent to specific problems, which may degrade the performances of classifiers in terms of prediction accuracy, sensitivity, specificity, and recall rate. The main goal of a multi-objective optimization problem is to identify the subsets of the given features. To this end, a hybrid cat swarm optimization (HCSO) algorithm is proposed in our paper for performance improvement of the basic cat swarm optimization (CSO) that incorporates guided and competitive & inherent characteristics into the original CSO. The performance of HCSO has been tested by finding the optimal feature subset for 15 benchmark datasets. The number of class labels for these datasets vary between 2 to 40. The time complexity analysis of both CSO and HCSO has also been evaluated. Moreover, the performance of the proposed algorithm has been compared with that of simple CSO and other state-of-the-art techniques. The performances obtained by HCSO have an average 2.68% improvement with a standard deviation of 2.91. The maximum performance improvement is up to 10.09% in prediction accuracy. Tested on the same datasets, CSO has yielded improvements within the range of -7.27% to 8.51% with an average improvement 0.9% and standard deviation 3.96. The statistical tests carried out in the experiments prove that HCSO manifests a moderately better feature selection capacity than that of its counterparts.

Automated summary

The study introduces a hybrid cat swarm optimization (HCSO) algorithm, which shows better performance in feature subset selection compared to the traditional CSO through testing.