<p>Current research in emotion recognition using physiological signals faces challenges such as handling high-dimensional data and managing inter- and intra-individual differences across subjects. To address these challenges, we propose a novel preprocessing approach for EEG signals combined with a stack Graph Neural Network (stackGNN) to enhance generalization in emotion prediction across different subjects. Initially, the multi-channel EEG signals are transformed into various graph-type data representations, which consolidate information from multiple EEG channels into single-node representations within the output graph. We then employ diverse graph neural network architectures to construct a stackGNN model. The model first feeds the graph-transformed EEG data into various graph convolutional layers, extracting a wide range of features from the signal. These features are then integrated through a feature aggregation layer, providing initial probability predictions for each graph-type model. Finally, a multilayer perceptron network refines these predictions to generate the final output. We validate our approach using the public DEAP dataset, where experimental results demonstrate that our method outperforms other approaches with improvement in both accuracy (up to 97.18% and 67.55% for subject-dependent and subject-independent tasks, respectively), and consistency.</p>