Dimensionality reduction and attribute selection using swarm based algorithm for dna microarray

Abstract

With the increase in the real-world data, handling the data during the implementation of optimization and machine learning applications is challenging. The high dimensional data is comprised of the instances along with the attributes belonging to the corresponding classes. The presence of a smaller number of instances with large number of attributes results in the curse of dimensionality. Redundant and irrelevant attributes are present in the data which decreases the performance of the classification algorithm. Attribute selection plays a vital role in optimization and machine learning that involves huge datasets. Classification accuracy of any learning model depends on the dimensionality of data and attributes selected. This leads to a multi-objective problem of obtaining high classification accuracy with fewer attributes. In this research work, a multi-objective optimization algorithm with greedy crossover for attribute selection and classification is proposed. A wrapper based Binary Bat Algorithm (BBA) with Support Vector Machine (SVM) as evaluator is implemented for attribute selection. In general, the optimization algorithms have the tendency to prematurely converge with sub-optimal solutions. This reduces the quality of the attribute selected and efficiency of the algorithm. Here, a multi-objective binary bat algorithm with greedy crossover is proposed to reset the sub optimal solutions that are obtained due to the premature convergence. The evaluation of the attributes selected is done using the Support Vector Machine with 10-fold cross-validation. The proposed algorithm is implemented and evaluated with the benchmark datasets available in the UC Irvine (UCI) repository. Classification accuracy of 89.25%, 96.45%, 96.57% and 88.50% using the Australian, Ionosphere, Wisconsin Breast iv Cancer (Original dataset) and Musk is obtained. The proposed multi objective binary bat algorithm with greedy crossover yields better performance over the existing algorithms. In the world of digitization and advan