Strategies for Class Imbalance in Emotion Analysis: A Comparative Evaluation

Emotion analysis is vital in applications like social media monitoring, marketing, advertising, entertainment and mental health assessment. A major challenge is class imbalance, where the most common labels are called majority classes and the least common labels are called minority classes, leading to biased models. This study explores methods to address class imbalance in emotion datasets, including cost-sensitive learning, adaptive thresholding, semi-supervised learning, and data augmentation, using the SemEval 2018 E-c dataset. In this study, the proposed model combines all of these techniques, resulting in an improvement of $\mathbf{3 5. 3 1 \%}$ in Macro, $\mathbf{3. 2 7 \%}$ in Micro and 12.97% in Weighted metrics. This highlights the importance of class balancing strategies in emotion analysis for more robust detection systems.