Emotion recognition can be used in clinical and nonclinical situations. Despite previous works which mostly used time and frequency features of electroencephalogram (EEG) signals in subject-dependent emotion recognition issues, we used multiscale fuzzy entropy as a nonlinear dynamic feature. The EEG signals of the well-known Database for Emotion Analysis Using Physiological signals dataset was used for classification of two and three levels of emotions in arousal and valence space. The compound feature selection with a cost of average accuracy of support vector machine classifier was used to reduce feature dimensions. For subject-dependent systems, the proposed method is superior in comparison to previous works with 90.81 % and 90.53 % accuracies in two-level classification and 79.83 % and 77.80 % accuracies in three-level classification in arousal and valence dimensions, respectively.