A feature-optimized machine learning system for real-time crisis tweet identification

Preparing and organising a relief effort using useful social media content in a crisis situation is crucial. Twitter, Facebook, LinkedIn, and many more social media platform produces immense data. Practical implementation for necessary automatic sorting and extracting specific text by classifying tweets into useful information, which helps humanitarian needs. Demanding less computational resources to concentrate on text models to scale the work from the CrisisMMD dataset. Machine Learning (ML) classifier test for converting text by using Term Frequency Inverse Document Frequency (TF-IDF), Word2Vec, and Bag of Words (BoW). Performance measure is used to evaluate accuracy, macro-averaged precision, recall, F1-score, and ROC-AUC, and tested for statistical significance and ROC curves analysis, with standard metrics on the classifiers Logistic Regression (LR), Support Vector Machines (SVM), Multinomial Naïve Bayes (MNB), Decision Trees (DT), Random Forests (RF), XGBoost, and a combined ensemble model. Better results were obtained with the Word2Vec model than with TF-IDF, and BoW consistently delivers effective word-count-based methods, indicating simpler methods. Logistic Regression (LR) and SVMs achieve consistent and reliable results in linear models when used in an ensemble approach with measured criteria yielding the most well-rounded performance. To identify humanitarian needs and class distribution, and to correct the models to strengthen especially critical disaster response. The research focuses on demonstrating traditional ML techniques, paired with reliable test filtering and the right extraction features. Disaster-related test filtering to present real-time monitoring supports humanitarian decision-making.