Retweet Prediction on Earthquake Tweets

This study aims to predict retweet engagement levels for tweets posted after the February 6, 2023, Kahramanmaraş earthquakes in Turkey. A dataset of 4.49 million tweets was collected and preprocessed by removing nontextual elements, applying Turkish spell checkers, lemmatization, and eliminating stop words. Features such as ‘Like’, ‘Retweet’, and ‘Reply’ and new combinations of those features were created. By analyzing these features and the generated features, it was found out which of them is the most suitable class separator. Tweets were categorized into two classes based on retweet interactions: “moderate-high” and “non-low”. Feature selection involved vectorizing text using inverse document frequency (IDF) values, with 400 words selected for the dataset. The ‘Retweet’ feature with a threshold of 2 was identified as the best class separator. Four models named as Random Forest, Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and DistilBERTurk were evaluated. Through the utilization of these models, we concurrently conducted a comparative analysis of the performance between manual vectorization and automatic vectorization techniques. XGBoost performed best, achieving 0.7340 accuracy, a ROC-AUC score of 0.81, and high precision and recall metrics for both interaction classes.