Feature selection using a genetic algorithm-based hybrid approach

Abstract

The present work proposes a hybrid feature selection model aimed at reducing training time whilst maintaining classification accuracy. The model includes adjusting a decision tree for producing feature subsets. Such subsets’ statistical relevance was evaluated from their resulting classification error. Evaluation involved using the k-nearest neighbors’ rule. Dimension reduction techniques usually assume an element of error; however, the hybrid selection model was tuned by means of genetic algorithms in this work. They simultaneously minimise the number of features and training error. Contrasting with conventional methods, this model also led to quantifying the relevance of each training set’s features. The model was tested on speech signals (hypernasality classification) and ECG identification (ischemic cardiopathy). In the case of speech signals, the database consisted of 90 children (45 recordings per sample); the ECG database had 100 electrocardiograph records (50 recordings per sample). Results showed average reduction rates of up to 88%, classification error being less than 6%.