Analyzing Tweets for Disaster Prediction

The application of Natural Language Processing (NLP) in analyzing disaster-related tweets toward enhancing crisis management efficiency is an important machine learning problem. Utilizing a dataset of 7,613 disaster related tweets, this study conducts a comparative analysis of three contemporary machine learners: XGBoost, CATBoost, and Random Forest. The research aims to determine which model most effectively classifies tweets as belonging to the disaster and non-disaster classes. In pursuit of this goal, the dataset underwent substantial preprocessing, including text normalization, stop word removal, and lemmatization. Feature extraction was performed using Count Vectorizer, tailoring the dataset for machine learning. The empirical methodology encompassed a detailed configuration and hyperparameter tuning of each model, followed by an implementation of a majority-vote ensemble to potentially enhance accuracy and generalizability of the predictions. The models were evaluated through a stratified five-fold cross-validation, focusing on accuracy, precision, recall, and F1-score as the performance metrics. Our case study results indicate slight differences in the performance of the individual models, but overall, the models were comparable in performance with the ensemble learner approach, demonstrating effective performance and generalization. The study provides insights into the efficacy of different machine learning models in processing disaster-related social media data and underscores the potential of ensemble techniques in NLP tasks. These findings have practical implications for enhancing real-time disaster response and management strategies. The research contributes to the broader understanding of NLP applications in crisis and disaster related data.