Fake News Detection

Fake news spreads fast, and telling it apart from real reporting by eye is getting harder. This project uses supervised machine learning to classify news articles as REAL or FAKE based on their text content.

It follows the classic approach popularized by the DataFlair "Detecting Fake News" tutorial: TF-IDF vectorization combined with a PassiveAggressiveClassifier.

How it works

1. TF-IDF Vectorization — converts article text into numeric features weighted by how important each word is to a document relative to the whole corpus.
2. PassiveAggressiveClassifier — an online-learning linear classifier that is fast to train and performs well on high-dimensional sparse text data.
3. Evaluation — accuracy score, confusion matrix, and a classification report (precision/recall/F1) on a held-out test set.

Project structure

Fake News Detection/
├── README.md
├── requirements.txt
├── data/
│   └── news.csv              # sample dataset (see note below)
├── model/
│   ├── train.py              # standalone training script
│   ├── vectorizer.pkl         # saved TF-IDF vectorizer (after training)
│   └── classifier.pkl         # saved classifier (after training)
└── docs/
    └── Fake_News_Detection.ipynb   # step-by-step Jupyter notebook walkthrough

About the dataset

data/news.csv currently contains a small, synthetic sample (40 rows) so the whole pipeline can be run immediately and end-to-end as a demo. It was generated for this project and is not a real news dataset.

The dataset link originally referenced for this project points to a Google Drive file (news.zip) that requires an authenticated browser session to download, so it could not be fetched automatically here. For real results, swap in a full-size dataset such as the ~6,300-article Kaggle "Fake News detection" dataset used in the original DataFlair tutorial (or any similar dataset), formatted as a CSV with at least these columns:

column	description
title	headline of the article
text	full body text of the article
label	REAL or FAKE

Just drop your CSV in as data/news.csv (same column names) and everything else works unchanged — or pass a different path/column names via command-line flags (see below).

Setup

cd "Fake News Detection"
pip install -r requirements.txt

Usage

Option 1: Run the training script

cd model
python train.py --data ../data/news.csv

Optional flags:

• --text-col — name of the text column (default: text)
• --label-col — name of the label column (default: label)
• --test-size — fraction of data held out for testing (default: 0.2)
• --out-dir — where to save vectorizer.pkl / classifier.pkl (default: current dir)

This prints accuracy, a confusion matrix, and a classification report, then saves the trained vectorizer and classifier as .pkl files.

Option 2: Explore the Jupyter notebook

jupyter notebook docs/Fake_News_Detection.ipynb

The notebook walks through every step interactively — loading data, TF-IDF vectorization, training, evaluation (with a confusion matrix heatmap), and testing the model on a custom headline — with explanations along the way.

Making a prediction with a saved model

import pickle

with open("model/vectorizer.pkl", "rb") as f:
    vectorizer = pickle.load(f)
with open("model/classifier.pkl", "rb") as f:
    classifier = pickle.load(f)

text = "Your article text goes here..."
prediction = classifier.predict(vectorizer.transform([text]))
print(prediction[0])  # 'REAL' or 'FAKE'

Notes & limitations

• Accuracy numbers from the bundled sample dataset are not meaningful — the sample is tiny and only exists to prove the pipeline runs. Retrain on a real, larger dataset before drawing any conclusions.
• This is a text-style classifier, not a fact-checker: it learns patterns in wording, tone, and structure that correlate with the labels in its training data. It cannot verify real-world facts, and its accuracy is only as good as the dataset and labels it's trained on.
• You may see a FutureWarning about PassiveAggressiveClassifier being deprecated in newer versions of scikit-learn (in favor of SGDClassifier with specific parameters) — this is harmless for current versions and doesn't affect results.

Possible extensions

• Compare against other models: Logistic Regression, Multinomial Naive Bayes, SVM.
• Add cross-validation and hyperparameter tuning (GridSearchCV).
• Combine title and text as input, or engineer additional features (article length, punctuation patterns, source metadata).
• Add explainability — inspect which TF-IDF terms most influence each prediction.
• Build a simple web UI (e.g. Flask/Streamlit) around the saved model for live predictions.

Issues

• If you found any errors, please make an issue report; really appreciate it 🤭