Multiple-Instance Lasso Regularization Via Embedded Instance Selection For Emotion Recognition

Since emotions affect physical and psychologically the health of people, their identification is crucial for understanding human behavior. Despite the several systems developed in this regard, most of them underperform on people with disabilities, their setup is sensitive to noise or non-emotional stimuli. Recent studies consider electroencephalographic (EEG) signals for understanding emotional responses due to reflecting the activity of the central nervous system. However, the non-stationary nature of EEG signals demand elaborated signal processing approaches because not all time instants hold information related to the stimulus-response. This work proposes a temporal analysis approach, termed MILRES, based on the Multi-Instance Learning framework that includes a multiple instance Regularization with LASSO penalty and an Embedded instance Selection. We test MILRES in discriminating two states (high and low) of the valence and arousal emotional dimensions from the DEAP dataset. The proposed approach reaches 84.4% accuracy and 79.5% F1-score for valence, and 81.9% accuracy 67.9% for arousal. Such results evidence that MILRES outperforms other EEG-based emotion recognition approaches from the state-of-the-art, with the additional benefit of identifying the brain areas involved in processing emotions.