Inertia factor and crossover strategy based particle swarm optimization for feature selection in emotion classification

Abstract

Emotion recognition using electroencephalography (EEG) is a better choice because it can’t be easily mimicked like facial expressions or speech signals. The emotion of EEG signals is not the same and vary from human to human, as everyone has different emotional responses to similar stimuli. Existing research has achieved lesser classification accuracy as it relies on whole feature subsets that include irrelevant features for classifying emotions. This research proposes the inertia factor and crossover strategy (IFCS)-based particle swarm optimization (PSO) algorithm to select relevant features for classification, which removes irrelevant features and enhances classification performance. Then, the self-attention with gated recurrent unit (SA-GRU) method is developed to classify the valence and arousal emotion classes, which focuses much on the significant parts of emotions and reaches high classification accuracy. The proposed IFCS-PSO and SA with GRU method achieved an accuracy of 98.79% for the valence class and 98.03% for the arousal class of the DEAP dataset, outperforming traditional approaches such as convolutional neural networks (CNN).