Facial Expression Recognition (FER) has emerged as a pivotal research domain within computer vision and affective computing. Despite technological advances in deep learning architectures such as CNNs, VGG-16, ResNet-50, and MobileNet, FER systems continue to face critical challenges, particularly data imbalance where some emotional categories are abundantly represented while others appear far less frequently. This imbalance leads to biased model performance and poor recognition rates for minority emotions, which are often the most crucial in sensitive applications. While numerous systematic reviews have examined fundamental FER aspects including recognition techniques, deep learning methodologies, and general challenges, there is a notable absence of focused systematic literature reviews that comprehensively analyze imbalance-specific solutions within FER. Our analysis revealed four primary approaches to handling data imbalance in FER: 1) Loss Function approaches - the most prevalent due to their simplicity and computational efficiency, 2) Generative Network approaches demonstrating the highest performance gains through sophisticated synthetic data generation; 3) Resampling approaches offering intuitive solutions through oversampling and undersampling techniques; and 4) Learning approaches employing multi-stage and ensemble architectures for sophisticated representation learning. © 2013 IEEE.