2.1.3. Normalization

EEG signals vary widely in amplitude depending on age, sex, and other factors like changes in subjects' alertness during the day. Hence, it is necessary to normalize measured values to deal with this variability.

There are three possible approaches to normalization. The first is to record reference conditions without stimulus on the subject. The values obtained can be normalized by subtracting the reference value, then dividing by the reference value (or subtracting the reference value), and then dividing by that same value. The second approach also requires reference conditions. Those values are included in the feature vector, which will have twice the characteristics that make up the "baseline matrix". The third approach normalizes the data separately by obtaining a specific range, for example, between −1 and 1. This method applied to each feature independently ensures that all characteristics have the same value ranges [38,39].

The effect of normalization and its influence on the entire process of emotion recognition is not ye<sup>t</sup> evident. However, some studies show that normalization allows the characteristics to be generalized so that they can be used in cross-subject emotion recognition. Tangentially, data normalization helps machine learning algorithms' efficiency due to faster convergence.
