A Reproducible Hybrid Machine-Learning Workbench For Fake News Detection on LIAR And Fake News Net

Abstract

Fake-news detection requires not only accurate classifiers but also transparent, reproducible, and extensible experimental workflows. This paper presents VeritasLab, a complete research workbench for fake-news detection using the LIAR and FakeNewsNet datasets. The system implements TF-IDF with Linear Support Vector Machines, BiLSTM with learned attention, and fine-tuned BERT-family transformers. The workbench provides data exploration, preprocessing controls, held-out evaluation, bootstrap confidence intervals, confusion matrices, model comparison, local explainability, model persistence, and research-result export.