Exploring the Time-efficient Evolutionary-based Feature Selection Algorithms for Speech Data under Stressful Work Condition

Abstract

Initially, the goal of Machine Learning (ML) advancements is faster computation time and lower computation resources, while the curse of dimensionality burdens both computation time and resource. This paper describes the benefits of the Feature Selection Algorithms (FSA) for speech data under workload stress. FSA contributes to reducing both data dimension and computation time and simultaneously retains the speech information. We chose to use the robust Evolutionary Algorithm, Harmony Search, Principal Component Analysis, Genetic Algorithm, Particle Swarm Optimization, Ant Colony Optimization, and Bee Colony Optimization, which are then to be evaluated using the hierarchical machine learning models. These FSAs are explored with the conversational workload stress data of a Customer Service hotline, which has daily complaints that trigger stress in speaking. Furthermore, we employed precisely 223 acoustic-based features. Using Random Forest, our evaluation result showed computation time had improved 3.6 faster than the original 223 features employed. Evaluation using Support Vector Machine beat the record with 0.001 seconds of computation time.