Emotion recognition from EEG signals represents a significant area of research within cognitive computing. Key challenges in this domain include the extraction of meaningful and discriminative features from non-stationary EEG signals and the development of highly accurate predictive models. To address these challenges, artificial intelligence (AI) techniques have been widely adopted to construct decision-support systems across various specialized domains. In particular, recent advancements in brain-computer interface (BCI) research have seen increased application of deep learning (DL) models for the interpretation and analysis of neural data. However, the inherent complexity and black-box nature of deep learning architectures raise concerns regarding the interpretability and transparency of model predictions. To enhance the interpretability of AI algorithms and their decision-making processes, an interpretable lightweight EEG-based model serves as a promising solution. This study proposes a novel attention-based deep learning architecture combined with an interpretable artificial intelligence technique for determining the significance of spatial-temporal features and the model’s predicted output for classifying human emotional states based on EEG signals. The publicly available DEAP emotion dataset is evaluated using cross-validation for multi-label classification across four primary emotional dimensions—arousal, valence, dominance, and liking–achieving classification accuracies of 73.83%, 79.29%, 62.13% and 59.37%, respectively. The interpretable AI saliency map highlights the most significant brain regions associated with each emotional state at the time of prediction. The experimental results demonstrate that the proposed framework outperforms existing methods in terms of prediction accuracy approx 3%. Furthermore, the proposed method achieves high classification performance while ensuring interpretability for emotion recognition applications.