Multimodal Dataset for Personality Prediction

ABSTRACT: Personality strongly influences human thoughts, emotions, and behaviors. However, existing datasets rarely integrate multi-modal neurophysiological signals for personality prediction. To address this gap, we present a multi-modal dataset that systematically combines four kinds of signals recorded during participants' viewing of different emotion videos. One hundred and fourteen participants viewed emotional video clips while their electroencephalograp (EEG), galvanic skin response (GSR), photoplethysmography (PPG), and facial videos were recorded, and their personality traits were measured through the BFI-44 questionnaire. Subjective affective states were assessed using short positive affect and negative affect schedules (PANAS), valence-arousal-dominance (VAD), and seven discrete emotion ratings. Technical validation confirmed the effectiveness of emotion induction and the quality of the collected signals. Extensive experiments using different modality combinations for personality prediction were conducted, demonstrating the dataset's utility and reliability for personality prediction. Four classifiers (kNN, SVM, RF, and MLP) were tested and RF models achieve up to approximately 90\% average accuracy in binary classification of Big Five traits across induced emotions. This represents, to our knowledge, the first publicly available dataset linking Big Five personality traits, subjective affect, and multi-modal neurophysiology under controlled emotional stimulation.