Abstract
The proliferation of misinformation on social media platforms like Instagram poses a significant threat, eroding public trust and potentially endangering public health. While research in this area has grown considerably, a notable gap remains in addressing misinformation in languages other than English. This study bridges that gap by establishing a comprehensive performance benchmark for detecting misinformation on Instagram, focusing specifically on Persian-language content related to COVID-19. To support this research, we constructed a novel labelled dataset of 27,000 Persian comments collected from Instagram. We employed five established machine learning and deep learning approaches for misinformation analysis: XGBoost (with TF-IDF embedding), LSTM (with GloVe embedding), CNN, KNN and BERT. To prevent overfitting, we conducted model training and validation, evaluating performance using confusion matrices. The results demonstrate that the LSTM model achieved the best classification performance, with an accuracy of 0.97, precision of 0.91, recall of 0.85 and an F1-score of 0.85. This outcome establishes the first high-performance baseline for deep learning methods on this unique Persian misinformation dataset. Furthermore, the training and validation curves for the LSTM approach indicate its effectiveness in mitigating overfitting.