Traditional Machine Learning-Based Fake News Detection System: Complete Implementation from TF-IDF to Logistic Regression

This article introduces a fake news classification system built using traditional machine learning techniques, with core components including TF-IDF feature extraction and logistic regression models. It implements a complete workflow from data preprocessing to web application deployment. The system performs efficiently in resource-constrained scenarios or those requiring high interpretability, providing a lightweight solution for fake news detection.

The spread of fake news in the digital age has become a social problem, and manual review is time-consuming and costly. This project chooses traditional machine learning methods to demonstrate how to build an efficient detection system in resource-constrained scenarios or those requiring interpretability, making up for the shortcomings of deep learning solutions in these aspects.

A binary dataset (Fake.csv and True.csv) containing real and fake news is used, with each entry including fields such as title and body text. Preprocessing steps: text lowercasing, removal of URLs/punctuation/special characters, merging of title and body text (since titles contain core information), to ensure data quality.

TF-IDF is used to convert text into numerical features (term frequency + inverse document frequency), using scikit-learn's TfidfVectorizer (including stopword filtering and N-grams). Logistic regression is selected as the model (suitable for high-dimensional sparse features and highly interpretable), with Naive Bayes as the baseline, and 5-fold cross-validation is used to ensure generalization ability.

Accuracy, precision, recall, and F1 score (primary metric) are used to evaluate the model. To provide statistical confidence, Bootstrap resampling is used to estimate the confidence interval of the F1 score, quantifying the reliability of the model's performance.

An interactive Streamlit web application is developed, supporting functions such as real-time prediction and performance display. Deployment methods: local (streamlit run app.py) or cloud (Streamlit Community Cloud). The code is modularly designed (data_utils, text_preprocessing, etc.), and TF-IDF is integrated into the Pipeline to avoid data leakage.

Current limitations: The dataset may have topic bias, only supports English, and TF-IDF ignores semantic relationships. Improvement directions: Introduce external knowledge bases, explore ensemble learning, build multi-source datasets, and develop domain-specific models.

This project demonstrates the effectiveness of traditional machine learning in fake news detection. The full workflow implementation (from data to deployment) provides a practical case for beginners, and the lightweight solution is suitable for scenarios requiring interpretability. Fake news detection needs continuous evolution, and this project lays the foundation for subsequent research.