Analysis and visualization of EEG data towards academic emotion recognition

• Main Author: Azcarraga, Judith Jumig
• Format: text
• Language: English
• Published: Animo Repository 2014
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/378
• Institution: De La Salle University

Brainwaves are analyzed, visualized, and are used to predict academic emotion based computational intelligence techniques, namely decision trees, Multi-Layered Perceptrons (MLP), k-means clustering, and Self-Organizing Maps (SOM). The brainwaves or EEG signals were collected from fifty six (56) academically-gifted students using an Emotiv EPOC sensor mounted on their head while they learn independently using two computer-based learning systems. Decision trees, as the baseline classifier, are able to classify the four academic emotions, namely frustrated, confused, bored, and interested, with prediction rates of only around 0.50. Multi-Layered Perceptrons are thus used to enhance the performance rates, achieving rates of 0.52 to 0.64 for Accuracy and 0.52 to 0.65 for Area Under the Curve (AUC). Prediction performance is also further improved by isolating the datasets based on gender, and/or handedness, and building separate classification models for each. Once separate classification models are built for the restricted sets, accuracy and AUC rates as high as 0.75 are achieved for datasets where all learners are only male and right-handed. Moreover, performance rates are shown to further improve when selective prediction is performed, i.e. only when the number of very high or very low feature values among the instances is high, such as when the number exceeds 20% of the 126 features used. As a further step in the analysis and visualization of the EEG data, unsupervised clustering is performed using k-means clustering and SOM. Using these two clustering methods, a new statistics-based method is proposed for determining the influence of the learners profile, such as gender and handedness, on the formation of the clusters. Experimental results on the use of a novel structure for a SOM, referred to as structured SOM, are also presented. These experiments include a newly developed measure for quantifying the quality of a trained structured SOM. In addition to the new methods for the study of the k-means clusters, and the work on structured SOM, the detailed discussion on the manner by which the EEG data can be properly cleaned, prepared and pre-processed, including several methods for feature selection and dimensionality reduction, provide very useful information for future in-depth studies on EEG data and how they should be used to predict the academic emotions of learners.