<![CDATA[User reviews for the Instagram application represent a vast and unstructured data source, offering valuable insights into public sentiment. However, their sheer volume and informal linguistic nature necessitate efficient and accurate automated analysis. This study conducts a comprehensive comparative analysis of three prominent Transformer-based models: BERT-base-uncased, DistilBERT, and BERTweet, to determine the most optimal architecture for sentiment classification of these reviews. Utilizing a dataset of 30,000 preprocessed reviews sourced from Kaggle, the models were fine-tuned for a binary classification task (positive/negative). Performance was systematically evaluated using accuracy, precision, recall, and F1-Score. The experimental results reveal a highly competitive performance landscape. BERT-base-uncased achieved the highest F1-Score (0.8475), establishing it as the best choice for balanced performance. Conversely, BERTweet, pre-trained on social media text, excelled in precision (0.8619), making it superior for reliable positive predictions. Meanwhile, DistilBERT demonstrated its value by offering a compelling balance of high performance (F1-Score 0.8373) and significant computational efficiency. This research concludes that the selection of an optimal model is not absolute but is contingent on specific application requirements, such as the priority of balanced accuracy, predictive reliability, or resource efficiency.]]>
Copyrights © 2026
