The rapid spread of fake news through digital and social media platforms poses significant social, political, and economic threats. This study presents a comprehensive framework for fake news detection that extends beyond binary classification to a multi-class approach using the LIAR benchmark dataset. The dataset categorizes political statements into six nuanced truthfulness levels, and our approach integrates both linguistic features and contextual metadata such as speaker identity, political affiliation, and statement venue. We begin with an exploratory analysis enriched by visualizations—such as word clouds and attribute-based distributions—to uncover patterns influencing misinformation. A diverse set of machine learning and deep learning models is evaluated, including Naïve Bayes, Random Forest, K-Nearest Neighbors, SVM, Logistic Regression, LSTM, BiLSTM, and transformer-based models like BERT, RoBERTa, DistilBERT, and ALBERT. Interestingly, Naïve Bayes outperforms more complex models, achieving an accuracy of 99.91%, underscoring that simpler algorithms, when paired with effective preprocessing and metadata integration, can yield superior performance. This work contributes a novel, interpretable, and scalable approach to multi-class fake news detection with metadata integration in a comprehensive, interpretable way while most studies focus on binary classification of fake news. By combining rigorous preprocessing with metadata-aware modeling, the framework offers a reliable tool for textual misinformation analysis. While focused on text data, this approach lays the foundation for future research into multimodal fake news detection, offering a practical and extensible solution to address the evolving challenges of digital disinformation.