<p>Recently, electroencephalogram (EEG)-based brain-computer interfaces for emotion recognition have gained more attention in several applications such as education, medical diagnosis, etc. However, these systems face major optimization challenges, including determining optimal temporal windows for machine learning algorithms, identifying the most informative frequency bands for emotion recognition, and selecting appropriate feature sets and classifiers that maximize recognition accuracy. The contribution of this study is a comprehensive approach to determine the optimal window sizes for both classical and tree-based classifiers, identifying the most effective frequency bands for each classifier, and introducing novel statistical frequency features derived from Welch’s power spectral density method, along with statistical features from the time domain. In this process, a total of 1,120 features were extracted, and mutual information was applied as feature selection to identify the optimal feature subset. The proposed approach was validated on the pre-processed versions of four EEG datasets: DEAP, GAMEEMO, DREAMER, and AMIGOS, employing both binary and multi-class classification with a subject-dependent 5-fold cross-validation. The obtained results demonstrated the high efficiency of the proposed approach across different datasets; for example, on the DEAP dataset, our approach achieved accuracies of 80.78% for valence and 83.34% for arousal. When applied to the other datasets, the approach also achieved remarkable results, and the comparison of our approach with some state-of-the-art methods showed the superiority of the proposed approach.</p>