Discovery of Shared Feature Mapping for EEG-based Emotion Recognition by Multi-Task Learning Approach

 Investigations have revealed that human emotions are resulted from their internal neural operations. The feedback of each emotion, sent from the skull surface, can be received and processed as a signal. Brain signals can be received and recorded by the EEG setup. In recent years, researchers and scholars have utilized various methods to capture and pre-process the signal, feature selection, dimensionality reduction and classification of brain signals. But, the number and type of extracted features play key roles in classification. Since it is unknown which feature operates more effectively and due to the fact that the number of used features are typically high and dependent on person, reduction in number of features and improving the efficiency of the classifier have been focused by many researchers. The purpose of this article is to provide a multi-task learning method to reduce the dimension and achieve a common space of features that describes the feelings of different people well. To show the efficiency of the proposed method, three well-known datasets are used, i.e. DEAP, SEED and DREAMER. Experiments are performed in two forms. In the first experiment, each channel is investigated separately. Channels with high efficiency are selected. The second experiment is performed by considering channels related to different parts of the brain (frontal, occipital, left hemisphere, right hemisphere). In the first experiment, the highest efficiency is about 80% and in the second experiment it is about 84%. Experimental results have shown that the proposed method have a higher efficiency than other comparing methods.