Fake News Detection with SVM, 1D CNN & LSTM
Abstract
A comparative study of Support Vector Machine, 1D Convolutional Neural Network, and LSTM models for classifying real vs fake news articles, trained on a public news corpus with TF-IDF features and deep text embeddings.
Introduction
Fake news (made-up or misleading stories dressed up as real reporting) spreads fast on social media and news sites. It can sway opinions, stir conflict, and chip away at trust in honest journalism.
This project looks at a simple but important question: given the text of an article, can software tell whether it is likely real or fake? To answer it, a few well-known approaches are compared on the same collection of articles.
Problem Statement
Human fact-checkers simply cannot keep up with how much content appears online every day. Automated help is needed to flag suspicious articles early, before they spread widely. The main challenges are messy text, tactics that keep changing, and building something that still works on articles it has never seen before.
Objectives
- Compare a few proven text-classification approaches for telling fake from real news.
- Clean and prepare the article text so models can learn from it.
- Check how well each approach performs on articles held back for testing.
- Identify which approach fits this problem best: and where the limits are.
Dataset
The work is built around a collection of labelled real and fake articles. Depending on the client, this can be a public news corpus, licensed or private archives, articles gathered from trusted outlets, or a custom set assembled for the project. The text is cleaned and tidied, lowercasing, removing clutter, and standardising wording, before any modelling begins.
Methodology
The article text is cleaned and turned into a form models can read, then three approaches are trained on the same data: SVM (with TF-IDF features), a 1D CNN for pattern spotting, and an LSTM that reads text in sequence. Each is trained and tested the same way so the comparison is fair.
The value here is the clear, apples-to-apples comparison. Rather than betting on one method, the project weighs a lightweight, fast approach against richer models that pick up more subtle wording patterns, showing which style of approach is the most dependable footing for a real fake-news filter.