Emotion Analysis Based on LASSO

In this paper, a technique for analysis of emotions has been proposed considering subject specific features based on EEG signal by employing LASSO regularization. Earlier most of the work focused on emotion recognition, by extraction and analysis of various features based on EEG sub-bands, primarily aiming towards the detection of the best possible set of features. However, subject dependency aspects have been ignored in such feature set even though emotions are very much subjective. Therefore, subject specific feature selection is a priority and is beneficial for emotion recognition studies. Hence, this paper utilizes sparse models defined by LASSO regularization for its ability to perform subject specific feature selection in a high dimensional EEG data. Power Spectral Density (PSD) features were first extracted for each EEG sub-band and consequently LASSO was applied to extract the most relevant subject specific sub- bands. For experimental validation, in total 32 subjects from DEAP database are considered. Following this, the features were examined for their emotion recognition ability by k-NN classifier. The proposed model gave better performance and aided in extracting robust and reliable features to yield a classification accuracy of 80% for arousal and 76% for valence score. Such findings ascertain that subject specific features are significant for emotion analysis and recognition systems and incorporating such features can make the systems more robust and reliable.