Hybrid DistilBERT and Machine Learning Pipeline for Twitter Fake News Detection

Abstract

The proliferation of misinformation on microblogging platforms, particularly Twitter, poses severe threats to public discourse and societal trust. This paper presents a lightweight, end-to-end fake news detection framework that integrates transformer-based natural language processing with conventional machine learning classifiers. The proposed system employs distilbert-base-uncased to generate contextual, 768-dimensional tweet embeddings through mean-pooling of the final hidden states. These semantic representations are subsequently consumed by two lightweight classifiers, namely Logistic Regression and Random Forest, with model selection governed by the Precision-Recall Area under Curve (PR-AUC) score computed on a held-out validation set. A Tkinter-based desktop graphical user interface (GUI) operationalises the trained model, supporting both single-tweet inference and batch CSV prediction with confidence scores. Experimental evaluation on a structured, balanced dataset of 200 annotated tweets yields a test-set accuracy of 100%, with perfect precision, recall, and F1-score for both classes. The confusion matrix and precision-recall curve confirm zero misclassification across all test samples. These findings substantiate the suitability of transformer-driven semantic embeddings combined with interpretable classifiers for accurate, computationally efficient fake news detection on short-form social media text.