Effective tweets classification for disaster crisis based on ensemble of classifiers

Abstract

In the field of disaster management, social media analytics has gained significant recognition. Social media platforms, particularly Twitter, have become an invaluable source for disseminating information during disasters, offering real-time updates on events, crisis reports, and casualty information. However, the deluge of information on social media can also be overwhelming, with a substantial amount of irrelevant content. To address this challenge, researchers leverage machine learning (ML) classifiers to automatically categorize disaster-related tweets. However, ML classifiers, while being effective, also face issues such as overfitting and class imbalance. This study proposes an ensemble-based approach that integrates a variety of linguistic and word embedding features, including Parts-Of-Speech (POS), hashtags, Term Frequency-Inverse Document Frequency (TF-IDF), GloVe, Word2Vec, and BERT. A range of supervised learning algorithms like Decision Trees, Logistic Regression, Support Vector Machines, and Random Forests, were evaluated individually and as part of ensemble methods like AdaBoost, Bagging, and Random Subspace. The results show that combining TF-IDF with word embeddings and using the AdaBoost ensemble model yields superior performance, achieving a classification accuracy of 98.92%. This represents a notable improvement over the conventional standalone classifiers and highlights the advantage of ensemble methods in enhancing model robustness and minimizing overfitting. The proposed approach demonstrates not only high predictive capacity but also scalability for real-time tweet filtering during emergencies. In addition to demonstrating the efficacy of ensemble methods in disaster tweet classification, this study also provides valuable insights for improving social media-based crisis response. It also establishes a foundation for future research, particularly in multi-lingual and multi-disaster scenarios.