An enhanced computational model based on social spider optimisation algorithm for EEG- based emotion recognition

• Main Author: Al-Qammaz, Abdullah Yousef Awwad
• Format: Thesis
• Language: English; English; English
• Published: 2019
• Subjects: QA Mathematics
• Online Access: https://etd.uum.edu.my/9047/1/s900489_01.pdf; https://etd.uum.edu.my/9047/2/s900489_02.pdf; https://etd.uum.edu.my/9047/3/s900489_references.docx; https://etd.uum.edu.my/9047/

Emotion is an element of human communication psychology that influences logical behaviour. Recently, human emotions are recognised by analysing signals of brain via electroencephalogram (EEG), however most studies show poor accuracy results. This is due to the limitations of existing feature extraction, feature selection and classification methods to address complex, chaotic and non-stationary EEG signals. Therefore, this research aims to develop an improved computational model for human emotion recognition model based on EEG multi-dimensional signals. This research has enhanced emotion recognition model in three parts. First part is feature extraction by proposing nearest-neighbour Grubbs based Discrete Wavelet Packet Transform (DWPT), where outliers are detected by Grubbs test and replaced to its nearest neighbour signal. Second part involves the feature selection method through developing an Improved Social Spider Optimisation (ISSO). This method is enhanced by incorporating Particle Swarm Optimization (PSO) global search towards better solution of spider movement behaviour. Third part is concerned on the development of Eagle Strategy Social Spider Optimisation (ESSO) for tuning parameters of Least Square Support Vector Machine (LSSVM) to avoid local optima problem. In this research, the proposed model is tested on the pre-processed EEG data obtained from Database for Emotion Analysis Using Physiological (DEAP) data set. The data was split into two groups according to subjects which are Data 1 and Data 2. Results showed that the proposed model outperforms the existing model. The maximum valence and arousal accuracies based on Data 1 were 76.39% and 83.33% respectively. While, the maximum valence and arousal accuracies in Data 2 were 72.22% and 81.94% respectively. The proposed EEG-based emotion recognition model contributes to the growing needs of an intelligent Brain-Computer Interface (BCI) and can facilitate the development of healthcare applications.