Positive-Unlabeled Learning Method for Positive Emotion Recognition Using EEG technology

Abstract

Emotion is a reaction of the human brain to external events, and the study of emotion recognition has substantial practical applications. Therefore, accurately recognizing and understanding positive emotions across different populations is crucial. Traditional image recognition technology cannot effectively identify emotions in individuals with impaired facial muscle control, such as elderly people in nursing homes with Alzheimer’s disease and patients with facial nerve paralysis (Bell’s palsy). Consequently, many machine learning methods have been widely applied to emotion recognition based on electroencephalogram (EEG) signals in recent years. In cases where the number of samples is sufficient, powerful deep learning methods can achieve high performance in emotion recognition. However, obtaining a large amount of reliably labeled emotional EEG data is arduous. We introduce a Positive-Unlabeled (PU) learning method for classifying EEG signals into Positive and Non-Positive emotions using a binary classifier developed with minimal labeled data. This approach utilizes a small volume of labeled data containing only positive emotion signals, combined with unlabeled data that includes both classes, effectively reducing the dependency on extensive, reliably labeled EEG data. The best accuracy achieved by this method is 93.95%. Experimental results on the dataset demonstrate theeffectiveness of our approach.