Recognizing affective state patterns using regularized learning with nonlinear dynamical features of EEG

In the present work, we aim to classify human emotional states categorized based on the arousal-valence model [1] by applying logistic regression (the original and L1 — regularized LR model) to nonlinear features extracted from electroencephalographic (EEG) signals. Recurrence quantifica­tion analysis (RQA) [2] was employed to effectively capture the underlying dynamics behind the complex reactivity corresponding to affective phenomenon. A benchmark dataset, DEAP [3], was used for our two-fold objectives: (1) to investigate the suitability of RQA measures and regularized learning method for emotion recognition, and (2) to compare the performances as well as topographic patterns of important channels for classifying emotional states with previous studies. The results demonstrated that our proposed method with selected RQA measures has better performance (test accuracy = 75.7% and F1 score = 78.1% on average) comparing to previous studies, and Li-regularized model is less over-fitted comparing to the LR.