Classifying Virality in Indonesian Twitter: A Cost-Sensitive Deep Learning Approach

This study focuses on classifying tweet virality using Indonesian Twitter (now known as social media X) data by leveraging pre-trained language models, specifically IndoBERT and RoBERTa, combined with deep learning techniques. To address the challenge of imbalanced data, cost-sensitive learning was employed to adjust class weights. Four models were evaluated: IndoBERT with balanced and twice-balanced weights, and RoBERTa with balanced and twice-balanced weights. Results showed that both IndoBERT and RoBERTa achieved 94% accuracy, but IndoBERT with twice-balanced weights demonstrated higher precision (59%) than RoBERTa, which exhibited superior recall (70%). IndoBERT's sensitivity to weight adjustments allowed for better control over false positives, while RoBERTa's performance remained stable despite weight changes. The research highlights the trade-off between precision and recall, offering insights for optimizing virality prediction. Future work could incorporate additional data sources and ensemble models to improve performance further.