Emotion recognition has become a cornerstone of human-computer interaction (HCI), as it enables intelligent systems to respond dynamically to users’ affective states. Among the available modalities, electroencephalography (EEG) offers a non-invasive and reliable way of capturing neural activity associated with emotions. This work investigates EEG-based emotion classification using the MAHNOB-HCI dataset and applies a complete pipeline comprising signal preprocessing, feature extraction, and machine learning-based classification. Statistical, spectral, and wavelet-derived features were extracted, while ensemble classifiers such as Random Forest (RF) and XGBoost were employed for prediction. Different preprocessing strategies, including baseline correction and normalization, were examined to evaluate their influence on classification accuracy. Experimental results demonstrate that XGBoost consistently surpasses RF, particularly in recognizing valence and arousal states, with wavelet and power spectral density features contributing most to performance. The study underscores the effectiveness of EEG-driven emotion recognition and provides practical insights for affective computing and neurotechnology applications. Future research directions include improving subject-independent generalization and enhancing model robustness through the integration of larger datasets and deep learning approaches.