A Subject-Wise Computational Framework for Classifying Questionnaire-Derived Dark Triad Profiles from Task-Onset EEG: Comparing Handcrafted, ROCKET, and EEGNet Representations

Task-evoked EEG can reveal individual differences in cognitive processing, but it also creates a machine-learning challenge: many epochs are recorded from relatively few participants, and evaluation can be misleading if within-subject dependence is ignored. This study compared handcrafted, ROCKET, and EEGNet representations under subject-wise validation for classifying questionnaire-derived Dark Triad profiles from task-onset EEG. Dark Triad traits were assessed with the Dirty Dozen and clustered into four exploratory multivariate profiles using k-means. EEG was recorded during a visual speeded decision task, and epochs were extracted from −200 to 1000 ms around task onset. The final dataset included 1780 retained task-onset epochs from 30 participants. Three representations were evaluated under identical five-fold subject-wise cross-validation: XGBoost with handcrafted EEG features, XGBoost with ROCKET-derived time-series features, and compact EEGNet. All performance estimates were based only on predictions from held-out participants. The handcrafted model achieved balanced accuracy of 60.9%, whereas ROCKET and EEGNet improved performance to 74.3% and 76.2%, respectively, with only a modest difference between the waveform-based representations. A participant-level label-shuffling analysis of the out-of-fold predictions indicated that prediction–label alignment exceeded chance for all models. Across models, the mean probability assigned to the true cluster increased with psychometric cluster centrality, suggesting that borderline profiles were harder to classify. Signed channel-wise ablation of EEGNet suggested distributed model sensitivity, with the largest positive effects over posterior/parietal electrodes. The findings highlight the importance of participant-level validation, EEG signal representation, and psychometric label structure, while emphasizing the need for external validation in larger independent cohorts.