Feature Dimensionality Reduction using a Dual Level Metaheuristic Algorithm

Abstract

The need to reduce the curse of dimensionality has spurred active research in the field of pattern recognition and machine learning. High dimensional data culminates in redundant as well as discriminant features. A vast amount of related literature offered the use of pattern classification methods such as Neural Network, Adaboost, Principal Component Analysis and Support Vector Machine, which adopts metaheuristics to extract important features from a pre-labelled training set. However, most of these individual metaheuristics algorithm are time and memory consuming when used to project the dataset into a lower subspace for subsequent classification. These drawbacks are reduced through a combination of two solution approaches whereby the strengths of two metaheuristics are brought together in a dual level algorithm. This research presents a dual level metaheuristic algorithm for feature dimensionality reduction using ant colony optimization (ACO) algorithm at Level 1 and genetic algorithm (GA) at Level 2. The dimension of the features in a face dataset will be reduced by using ACO at Level 1 to extract the relevant facial features in the training dataset. The output obtained will be fed into GA at Level 2 to select the discriminant feature subset that becomes the optimized parameter used by support vector machine (SVM) for classification. The developed algorithm suggests that the combination of one or more nature inspired metaheuristic algorithm is crucial for obtaining a high performance optimizer for feature dimensionality reduction to solve the problem of classification.