EEG-based Emotion recognition is more insightful than visual or auditory inputs in understanding human emotions due to direct access to underlying neural activity. Graph Neural Networks (GNNs) efficiently model EEG signals but often struggle with class imbalance. In this paper, we propose a novel oversampling technique, SVM-Guided GraphSMOTE, that focuses on minority class support vectors using a trained SVM on Differential Entropy features, and interpolates them with those of their nearest neighbors in the embedding space to generate informative synthetic graphs, while preserving the original graph structure. The proposed approach has been tested on the DEAP dataset in subject-independent settings. The experimental results outperform existing methods reported in the literature across the emotional dimensions of valence and arousal.