Machine Learning Strategies to Improve Generalization in EEG-based Emotion Assessment: \\a Systematic Review

Emotions are our internal compass and play a primary role in learning, reasoning, decision-making processes, and communication between individuals. The Information and Communication Technology (ICT) sector's interest in emotions has grown tremendously in recent years, shaping the concept of affective computing, an emerging field aimed at monitoring and predicting emotions in order to improve human-computer interaction Cambria et al. (2017); for instance, the introduction of affective loops makes it possible to implement increasingly adaptive human-machine interfaces and virtual assistants tailored to users Saganowski et al. (2020), or the outputs of emotion monitoring systems, in the healthcare context, can be useful in the treatment of psychological disorders based on emotional deficits, in autism Feng et al. (2018), in the improvement of wellbeing Healy et al. (2018), and in stress containment Saganowski (2022). In particular, in this context, there is a growing interest in the literature for Brain-Computer Interface (BCI) systems based on EEG signals Torres et al. (2020). In fact, the number of annual scientific publications indexed on Scopus database on the topic of EEG-based emotion recognition shows an exponential growth trend (see Figure 1). A critical issue underlying the processing and classification of EEG signals is their inherent variability among different subjects or different acquisition times (i.e.