This study presents a comparative experimental evaluation of emotion recognition and classification tasks using electroencephalography (EEG) and peripheral physiological signals, integrating public benchmarks (SEED and DREAMER) with a newly collected multimodal data set in our laboratory. The work systematically compares classical and deep learning approaches for valence-based emotion classification following standardized preprocessing and normalization across EEG, ECG, PPG, and EDA modalities. Support Vector Machine (SVM), Random Forest (RF), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) models were implemented to examine spatial, temporal, and multimodal representations. CNN achieved the highest accuracy on SEED (90.22%), RF performed best on DREAMER (83.80%), and the in-lab data set produced 98.37% with RF, highlighting strong reproducibility within the subject under controlled conditions. The findings demonstrate that aligning the model architecture with signal characteristics significantly enhances affective decoding performance. In general, this work establishes a unified and reproducible framework for multimodal emotion recognition and underscores the importance of cross-dataset validation to advance robust and generalizable affective computing systems.