Binary Optimization Using Hybrid Owl Optimization For Biomarker Selection From Cancer Datasets

The enhancement of medical technology generates a massive amount of disease data. For accurate disease detection, feature subset selection plays an important role in solving different classification problems. The selection of a good feature subset enhances the accuracy of the machine learning model and reduces the training time. However, it becomes tricky and challenging to identify the optimal feature subset from the high-dimensional datasets. Different hybrid models are proposed by various researchers to deal with these types of issues. The Owl optimization algorithm proves its efficiency in selecting the optimal features. Keeping an eye on the performance of this stochastic optimization, a binary version of the hybrid Owl optimization algorithm with simulating annealing is proposed to detect the biomarker features from the cancer datasets. K nearest neighbors (KNN) are used as a wrapper to find the best biomarkers. Eight benchmark datasets were employed for the proposed model’s performance evaluation. The results show that the proposed model is significantly better than binary Owl optimization, and other counterparts using various performance matrices such as accuracy, selection of best biomarkers, and computational time. The proposed model Owl-SA performs with an accuracy of 100% in the case of CNS and lymphoma datasets. And it retains the accuracy of 99.14%, 97.16%, 98.41%, 97.94%, 98.93%,98.32% for breast, colon, ALL-AML, ovarian, SRBCT, and MLL respectively.

I. Introduction

Data processing plays a vital role in preparing a dataset for a machine learning model. Feature selection is used to remove irrelevant and redundant features to enhance efficacy. Generally, there are two classes of feature selection methods: filter and wrapper. Filter methods are independent of classifiers and perform according to the nature of the dataset (information gain, correlation, variance, etc.). The wrapper feature selection depends on the classifier to obtain maximum classification accuracy. So most of the researchers prefer the wrappers in comparison with the filters for good accuracy. However, filter methods are also preferred in some circumstances in a machine learning model, as they are computationally less expensive than wrapper methods. To obtain the advantages of both filtering and wrapping, hybrid feature selection techniques are adopted by the researchers. Wrapper methods use an evaluating algorithm to measure the quality of the biomarkers, such as decision trees (DT), support vector machines (SVM), naive Bayes (NB), KNN, artificial neural networks (ANN), and linear discriminant analysis (LDA). Different traditional methods, such as sequential search (SFS), are used for the early detection of cancer disease. Stagnation in local optima, nesting effect, and high cost [1]. The most common limitation of SFS is computation. To avoid nesting effect limitations, different floating methods such as sequential forward floating search (SFFS) and sequential backward floating search (SBFS) are jointly used. For high-dimensional datasets, floating methods also fail to achieve good accuracy. A new generation of search algorithms known as ‘‘metaheuristic algorithms’’ evolved from the revolution in feature selection. Continuous effort is being made to improve the evolutionary algorithm’s performance, for example, particle swarm optimization (PSO), ant colony optimization (ACO), and artificial bee colonies (ABC) are examples of genetic algorithms (GA) and swarm intelligence (SI). There were also new optimization algorithms developed, including grey wolf optimization (GWO), grasshopper optimization (GOA), butterfly optimization (BOA), ant lion optimization (ALO), whale optimization algorithm (WOA), and harris hawk optimization (HHO) [2]. There are two varieties of metaheuristic algorithms: single solution-based and population-size-based metaheuristics. Algorithms depend on the nature of the exploration phase and exploitation phase. In the optimization phase, only one solution is processed at a time. single solution-based, where multiple solutions can be processed at the same time [4]. A population size-based metaheuristic begins by generating solutions from the initial population size and then iteratively replacing the existing population [5]. In recent years, researchers have accepted the use of hybridized metaheuristic algorithms to solve feature selection problems [6]. The main aim of adopting such a hybrid model is to identify the best solutions in order to achieve high performance in solving problems by balancing both exploration and exploitation [3]. Most of the population-based metaheuristic algorithms are efficient during exploration. So different local search algorithms are taken into consideration to enhance the exploitation phase and identify the best solutions [7].