EEG ‐Based Emotion Recognition Using Optimized Logit Boost and Relief Feature Selection

This study proposed a robust framework for recognizing affective states in the valence–arousal space from electroencephalogram signals because the framework combined Relief‐based feature selection with LogitBoost classifier and used systematic hyperparameter optimization. The model delivered accuracies of 80.70% for valence and 82.30% for arousal on the DEAP dataset because the pipeline emphasized compact feature subsets and efficient boosting, and the model also reduced computation relative to complex deep networks that demand heavy resources. The Relief began with 140 features and emphasized right frontal and occipital attributes, with F8, O2, AF4, and F4 contributing most because these sensors captured prefrontal control and visual processing linked to emotion. The optimized hyperparameters diverged for the two affective dimensions because the learning rate η stabilized at 0.1 for valence and 0.2 for arousal, which indicated distinct convergence behaviors that reflected different neural substrates and signal statistics across dimensions. These differences suggested a principled path to specialized models that target individual emotional axes because dimension‐specific tuning improved stability and yielded better generalization. Together, focused selection and careful optimization offered a credible alternative to deep learning because the approach delivered accurate, adaptive, and compute‐efficient EEG emotion recognition suited to clinical and BCI use. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.