On the prediction models for brain signal-based emotion recognition

This paper analyzes the factors contributing to the performance gap between subject-dependent and subject-independent EEG-based emotion recognition systems. With this key objective, the experiments are conducted in this work using two benchmark datasets; DEAP and SEED across classical machine learning algorithms, KNN, Random Forest, Linear Discriminant Analysis and a deep learning model like CNN and EEGNet. In cross-subject evaluations, CNN achieved the highest accuracy of 84% on the SEED dataset. To interpret these outcomes, Explainable Artificial Intelligence (XAI) techniques using SHAP (SHapley Additive exPlanations) values are employed to identify the most influential EEG features and frequency bands. This analysis not only highlights the key contributors to model decisions but also exposes feature-level inconsistencies that explain the reduced performance in subject-independent scenarios. This diagnostic study provides valuable insights into the limitations of current approaches and highlights the importance of explainability for developing robust EEG-based emotion recognition systems.