Streamwise Roberta: Efficient Event Classification from Rapid Twitter Data Flows

Abstract

The rapid growth of user-generated content on Twitter has created massive, high-velocity streams of text related to events such as disasters, politics, protests, riots, and terrorism, making real-time analysis challenging. A significant amount of useful information often remains unanalyzed, leading to delayed or missed insights. To address this issue, the study follows a comprehensive NLP-based framework focused on efficiency and scalability. The workflow begins with systematic data collection and preprocessing, including text cleaning, normalization, and tokenization, to remove noise and ensure high-quality input. Lightweight Robustly Optimized Bidirectional Encoder Representations from Transformers Approach is then used for contextual feature extraction, providing rich semantic representations while keeping computational overhead low. These embeddings form the basis for building effective classification models. Several learning approaches, including Random Forest Classifier (RFC), greedy tree-based methods, and deep learning models such as Deep Neural Network (DNN) with Stochastic Gradient Descent (SGD), are evaluated for comparison. Model performance is analyzed across multiple event categories, including disaster, political, protest, riot, and terror events. Experimental results show that although multiple models perform well, Histogram Gradient Boosting (HGB) offers superior efficiency and scalability. Its histogram-based feature binning and greedy optimization significantly reduce training time and resource usage. Consequently, Histogram Gradient Boosting (HGB) is identified as the most effective model for real-time, large-scale Twitter event classification and social media monitoring.