Enhancing multilabel classification for unbalanced COVID-19 vaccination hesitancy tweets using ensemble learning.

Purpose: Vaccination plays a crucial role in public health strategies, particularly in combating contagious diseases like COVID-19. Despite developing and authorizing numerous vaccines, achieving widespread immunity remains a challenge. Understanding individuals' willingness to receive vaccines is essential, and recent studies have explored acceptance rates across various regions. However, vaccine hesitancy is still considered a significant global health concern, even after the pandemic's peak.

Methods: This article explores the topic of classification in social media networks, specifically in tweets, focusing on the challenges of class imbalance and multilabel multiclass tweet classification. It investigates the effectiveness of oversampling in addressing unbalanced datasets and proposes an ensemble-based framework comprising various BERT models fine-tuned on diverse text corpora.

Results: The proposed model was applied to a benchmark dataset containing 12 classes with a highly imbalanced distribution. Compared to recent literature, the model achieved significant improvements in performance metrics. We observed a 2% increase in micro, macro, and weighted average F1 scores, a 4% increase in accuracy, and a 3% increase in average Jaccard index when using the proposed framework on this unbalanced multi-labeled benchmark.

Conclusion: The study investigated fine-tuning BERT models for improved classification in tweets on unbalanced datasets related to public health. The proposed ensemble approach with oversampling techniques yielded significant improvements in multiple metrics compared to existing methods. This suggests a promising approach to analyzing vaccine discourse on social media and potentially aiding public health efforts.