A Novel Approach for Feature Selection Using Artificial Neural Networks and Particle Swarm Optimization

Abstract

Advancements in technology make the application domains to collect large amounts of data that is used for future decision making by applying various techniques such as classification, regression, etc. These techniques comprise machine learning algorithms. The running time complexity of these machine learning algorithms is determined by the number of instances and the total of all features or dimensions in the dataset. All the instances are generally required. But all the features in the dataset may not be related or useful in the machine learning task, for example, in classification. In such a case, the dimensionality of the dataset is reduced to optimize the running time complexity of the algorithms without compromising the accuracy. Reducing the dimensionality is achieved by extracting only those features which are more relevant. Hence feature extraction or dimensionality reduction plays a significant role in obtaining high accuracy of machine learning algorithms within the acceptable time complexity. The existing techniques for feature extraction are principal component analysis and by using correlation coefficients. But these techniques required more time complexity and some sort of manual decisions for selecting optimal features. To extract the most optimal features, a novel algorithm is proposed using an evolutionary optimization technique called particle swarm optimization (PSO). As a fitness function in PSO, an artificial neural network classifier pertaining to its efficient performance in classification tasks. The performance of the proposed algorithm is verified by five benchmark datasets from the UCI repository. The model is evaluated with different performance measures:accuracy, precision, recall, and the F-1 score. Research results have proven that algorithm proposed here reduces the number of features by improving the performance of the machine learning sub-system when compared to conventional algorithms.