ABSTRACT
In emotion recognition using EEG, it is not generally agreed upon how much time an EEG signal sequence must have in order to maximize precision and recall rates. To the best of our knowledge, there is not a systematic evaluation of effects on classifier performance related to EEG signal durations. The human factors related to attention decreasing and tiredness increasing have imposed difficulties to create EEG datasets containing a rich variation of signal samples. This project proposes an evaluation of three different EEG datasets (DEAP, MAHNOB, and STEED) each one mainly characterized by short, intermediate and long signal (or stimulus) durations. Statistical evaluation pointed out that for an EEG dataset to be well-suited for emotion recognition it should have two main characteristics: emotion stimulus data should be publicly available and evaluated by world-wide volunteers, and media stimulus should have duration long enough to affect the subjects. Our statistical analysis revealed that, at least for the considered datasets, signals with duration longer than 60 seconds allow better classification results.