Which Model Feels Better? A Comparison of Computational Approaches to Emotion Detection in Social Media with Imbalanced Data

Emotion detection in social media remains challenging, particularly in polarized public debates where emotional expression is often imbalanced across categories. Addressing this challenge, this study compares the performance of multiple computational approaches using a gold-standard dataset of tweets about an ongoing geopolitical conflict. The dataset reflects the authentic, skewed distribution of emotions observed in real-world online discourse. We evaluated lexicon-based methods, classical machine-learning classifiers, deep-learning architectures, transformer models in both fine-tuned and zero-shot configurations, and a zero-shot large language model to assess their effectiveness in capturing both frequent and less frequently expressed emotions. Across approaches, transformer models, especially those fine-tuned for contextual emotion recognition, demonstrated the strongest overall performance, with emotion-specific fine-tuning offering a particular advantage for detecting rare emotion categories. These findings emphasize the importance of evaluating emotion detection methods under realistic class imbalance and highlight both the comparative strengths and limitations of widely used modeling strategies in applied social media research. This study advances emotion analysis and computational social science by offering practical guidance for selecting appropriate emotion detection methods in complex, imbalanced social media contexts.