Dataset and setup

We conducted our experiments on the DEAP dataset [19], which is perhaps the most popular dataset for the analysis of human affective states. Relevant to our research, DEAP provides EEG signals (32 channels) of 32 participants while watching 40 one-minute long excerpts of music videos. Participants rated each video in terms of valence, arousal, like/dislike, dominance, and familiarity. DEAP includes both raw and preprocessed EEG signals. In our experiments, we use the latter to ease replication and comparisons against previous work. Preprocessing includes downsampling the 512 Hz original signal to 128 Hz, removing electrooculography artefacts, and applying a band-pass filter in the [4,45] Hz range.¹

We divided the one-minute brain signals into short temporally consecutive segments of 1, 2, or 4 s long, without overlap.² DEAP includes a 3-second long EEG signal previous to the stimulus. Following [16], the average of the three one-second segments of this “rest-state” signal results in 32-D mean vector per channel, which is subtracted, also channel-wise, to each 1-second segment of the EEG signal during the presentation of the visual stimulus. Then, each channel is separately scaled to have unitary maximum absolute value.

Each pair (v, s) of video v∈{1,…,40} and subject s∈{1,…,32} has a label ℓ∈[1,9] for each emotional dimension (valence, arousal, and dominance), which corresponds to the subjective self-reported annotation. In this work, we considered only the valence dimension, and binarized its values into “negative” (ℓ≤5) and “positive” (ℓ>5), in line with much of previous work  [22, 32, 33]. Therefore, we consider a 2-way (binary) classification problem. The binarized labels are used as ground-truth for model training and performance evaluation on the test samples. Each individual segment inherits the label from the (v, s) signal it belongs to.