An Interpretable Multiscale Temporal Residual Framework for Gender-Specific EEG-Based Emotion Recognition

Electroencephalogram (EEG) sensors are well-suited for analyzing emotion-related neural dynamics due to their high temporal resolution in capturing brain activity. However, gender-related variability presents a significant challenge, as differences in neural dynamics between male and female subjects can influence the robustness and generalization of the emotion recognition systems. This work proposes an interpretable multi-scale temporal framework for analyzing gender-specific EEG dynamics in emotion recognition. The proposed framework employs a multi-scale temporal residual network as its computational backbone, integrating multi-resolution convolutional processing with residual recurrent modeling to capture both short-term and long-range temporal patterns in EEG signals. Extensive evaluations are conducted on three benchmark datasets, DEAP, SEED, and DREAMER, under various subject-dependent and subject-independent evaluation protocols. To enhance interpretability, explainable AI (XAI) techniques are employed to identify spatial, temporal, and representation-level EEG patterns contributing to model predictions. Experimental results demonstrate that the proposed framework consistently achieves stable and generalizable EEG temporal representations within each benchmark dataset across multiple evaluation settings, while exhibiting strong and competitive classification performance relative to existing methods. The analysis further reveals systematic gender-specific EEG activation patterns, highlighting the influence of demographic variability on emotion recognition performance. These findings underscore the necessity of incorporating gender-aware modeling strategies to develop robust EEG-based emotion recognition systems suitable for practical applications.