**6. Conclusions**

Subject-independent models fail to consider that the appraisal of one's emotional state is strongly related to personal factors, such as one's circumstances [39]. Subject-dependent models aim to tackle

this weakness, but they do so at the cost of significantly increasing data collection needs [40,41]. This implies that they have to be individually trained for each user and hence cannot be used with previously unseen subjects.

In this paper, a mixed framework to support automatic emotion recognition is proposed. Unlike most typical subject-dependent modeling approaches, the method uses data from all users to build the model, and can be used to make predictions for previously unseen users in an adaptive way, increasing performance as more training data become available. We first show that the existence of an inherent subject-related component in the EEG signals is a major obstacle when attempting to build a user independent model that is simultaneously valid for all subjects. Then, we propose a subject-based normalization procedure that is able to reduce the magnitude of this component when using PSD features. This straightforward normalization procedure is not intended to be a solution to remove this component, but rather a demonstration of the potential benefits of reducing its magnitude. The removal of the subject-dependent component in the signal is indeed feature and problem dependent, and an optimum approach cannot be generalized at this stage. This implies that there is still room for improvement by designing other normalization methods that are more efficient at this task.

The impact of the proposed method goes beyond the construction of inter-subject models for emotion detection from EEG signals. First, the same principles can be exported to other sources of information other than EEG, e.g., physiological, audio, and video. Second, these principles are not limited to the particular problem of emotion recognition. On the contrary, the subject-related component is intrinsic to the signal, and it is present regardless of the problem context.

**Author Contributions:** All authors were involved at all stages of the research. The original idea was proposed in a group meeting and progressively developed by all authors, who have all equally contributed from the conceptualization to the paper writing.

**Funding:** This research was partly supported by the Spanish Ministry of Economy and Competitiveness through projects TIN2014-59641-C2-1-P, PGC2018-096463-B-I00 and RTI2018-097045-B-C21; and the Ramón y Cajal gran<sup>t</sup> RYC-2017-22101.

**Conflicts of Interest:** The authors declare no conflict of interest.
