2.1. DEAP Dataset

The DEAP dataset is composed of the EEG records and peripheral physiological signals of 32 participants, which were recorded as each watched 40 1-min-long excerpts of music videos, relating to the levels of valence, arousal, like/dislike, familiarity, and dominance reported by each participant. Firstly, the ratings came from an online self-assessment where 120 1-min extracts of music videos were rated by volunteers based on emotion classification variables (namely arousal, valence, and dominance) [25]. Secondly, the ratings came from the participants’ ratings on these emotion variables, face video, and physiological signals (including EEG) of an experiment where 32 volunteers watched a subset of 40 of the abovementioned videos. The official dataset also includes the YouTube links of the videos used and the pre-processed physiological data (down sampling, EOG removal, filtering, segmenting, etc.) in MATLAB and NumPy (Python) format.

DEAP dataset pre-processing:

The data were down sampled to 128 Hz.

EOG artefacts were removed.

A bandpass frequency filter from 4 to 45 Hz was applied.

The data was averaged to the common reference.

The EEG channels were reordered so that all the samples followed the same order.

The data was segmented into 60-s trials and a 3-s pre-trial baseline was removed.

The trials were reordered from presentation order to video (Experiment_id) order.

In our experiments, we use the provided pre-processed EEG data in NumPy format, as recommended by the dataset summary, since it is especially useful for testing classification and regression techniques without the need of processing the electrical signal first. This pre-processed signal contains 32 arrays corresponding to each of the 32 participants, with a shape of 40 × 40 × 8064. Each array contains data for each of the 40 videos/trials of the participant, signal data from 40 different channels (the first 32 of them being EEG signals and the remainder 8 being peripheral signals such as temperature and respiration) and 8064 EEG samples (63 s × 128 Hz). As we are not working with emotion in this work, the labels provided for the videos on the DEAP dataset were not used in this experiment. Instead, we retrieved the original videos from the URLs provided and performed an exhaustive classification of them on the studied VDEPs. Of the original 40 videos, 14 URLs pointed to videos that were taken down from YouTube as of the 4 June 2019, therefore, only 26 of the videos used in the original DEAP experiment were retrieved and classified. The generated dataset was updated on the 5 May 2021.