📰 Fake News Detection with NLP & Machine Learning

An end-to-end Natural Language Processing (NLP) text classification pipeline designed to detect misinformation, fabricated articles, and clickbait with high precision using the WELFake benchmark dataset.

📓 Open Notebook • 📊 Dataset • 🚀 Try Inference • Report Issue

📌 Executive Summary

The rapid spread of online misinformation, propaganda, and unverified digital reporting presents severe challenges for media credibility, public health, and democratic integrity. Automated fake news detection systems provide crucial first-line verification tools for news aggregators, content moderation platforms, and digital journalists.

This project delivers a complete, production-oriented NLP classification pipeline trained on the WELFake (Word Embedding-based Learning for Fake News Detection) Dataset. The system:

Performs Linguistic Profiling: Analyzes headline vs. body text characteristics, token distributions, and lexical patterns distinguishing fake and legitimate reporting.
Executes Noise-Resistant Cleaning: Cleans text using targeted regex pipelines (removing URLs, HTML markup, specialized symbols, and anomalous whitespace).
Builds High-Dimensional N-Gram Representations: Employs Scikit-Learn TfidfVectorizer capturing both single words and two-word contextual pairs (unigrams + bigrams) across a 50,000-token feature space.
Delivers Rapid, Accurate Inference: Achieves 93.53% test accuracy and 0.935 F1-score with sub-millisecond per-article classification latency.

🏆 Key Highlights & Performance

Important

Test Accuracy: 93.53% on 3,972 held-out news articles.
F1-Score: 0.935 balanced across both Fake (label=0) and Real (label=1) classes.
Zero-Latency Inference: Highly efficient sparse TF-IDF + Logistic Regression pipeline executes in under 2 milliseconds per document.
Dual-Channel Text Fusion: Merges headline semantics (title) with in-depth narrative context (text) to maximize contextual awareness.

📊 Dataset Overview & Profiling

The model is built using the WELFake Dataset, a public benchmark combining 72,134 news records aggregated from four major sources: Kaggle, McIntire, Reuters, and BuzzFeed.

Class Label Convention

0: Fake News (Fabricated stories, deceptive claims, misinformation)
1: Real News (Verified reporting from reputable news agencies such as Reuters)

Sample Subset & Preprocessing Stats

Attribute	Full Dataset	Processed Modeling Subset
Total Articles	72,134	20,000 (10,000 Fake / 10,000 Real)
After Null Dropping	71,537	19,856 (10,000 Fake / 9,856 Real)
Train Set (80%)	—	15,884 articles
Test Set (20% Holdout)	—	3,972 articles

Linguistic & Length Metrics (Post-Cleaning)

Metric	Min	25th Percentile	Median (50th)	75th Percentile	Max	Mean
Title Length (Chars)	4	62	73	89	286	76.99
Body Length (Chars)	1	1,414	2,435	4,081	115,372	3,283.63
Combined Words	0	243	414	687	21,284	558.61

🔬 NLP Pipeline Architecture

flowchart TD
    A[Raw News Article: Title + Body] --> B[Data Ingestion & Null Handling]
    B --> C[Linguistic Profiling & Length Analysis]
    C --> D[Stratified Train-Test Split: 80% / 20%]
    
    subgraph Preprocessing [Text Cleaning & Normalization Engine]
        D --> E[Lowercasing]
        E --> F[Strip URLs & Links: http/www]
        F --> G[Strip HTML Tags: regex]
        G --> H[Alphanumeric & Punctuation Filtering]
        H --> I[Whitespace Normalization]
    end
    
    subgraph Vectorization [TF-IDF Feature Space]
        I --> J[TfidfVectorizer]
        J --> K[N-Gram Range: 1, 2 unigrams & bigrams]
        J --> L[Max Vocabulary: 50,000 features]
        J --> M[Stop Words: English, min_df=2]
    end
    
    K --> N[Sparse Term Matrix]
    L --> N
    M --> N
    
    subgraph Classification [Scikit-Learn Pipeline]
        N --> O[Logistic Regression Classifier]
        O --> P[Probability Calibration]
    end
    
    P --> Q[Test Evaluation: Accuracy, Precision, Recall, F1]
    P --> R[Production Inference: predict_news API]

🧹 Text Preprocessing & Feature Extraction

1. Robust Regex Text Sanitization

Raw web articles frequently contain HTML fragments, tracking URLs, and irregular typographical characters. The clean_text() function normalizes input text:

def clean_text(s: str) -> str:
    if pd.isna(s):
        return ""
    s = str(s).lower()
    s = re.sub(r"http\S+|www\.\S+", " ", s)           # Remove URLs
    s = re.sub(r"<.*?>", " ", s)                         # Remove HTML tags
    s = re.sub(r"[^a-z0-9\s\.\,\!\?\-']", " ", s)  # Keep valid punctuation
    s = re.sub(r"\s+", " ", s).strip()                  # Collapse whitespace
    return s

2. TF-IDF Hyperparameters

ngram_range=(1, 2): Captures individual salient words (e.g., "conspiracy", "reuters") as well as key phrases (e.g., "breaking news", "white house").
max_features=50,000: Restricts the vocabulary to the top 50,000 most informative terms, ensuring compact memory usage.
stop_words='english': Filters non-informative functional words.
min_df=2: Prunes one-off typographical typos and noise.

📈 Model Evaluation & Performance

The pipeline was fitted on the training split (15,884 samples) and evaluated on an independent holdout test split (3,972 samples).

Classification Report (Holdout Test Set)

Class	Precision	Recall	F1-Score	Support
0 (Fake News)	0.9399	0.9310	0.9354	2,000
1 (Real News)	0.9307	0.9397	0.9352	1,972
Overall Accuracy	—	—	0.9353	3,972
Macro Average	0.9353	0.9353	0.9353	3,972
Weighted Average	0.9353	0.9353	0.9353	3,972

Confusion Matrix (Test Set)

                  Predicted Fake (0)    Predicted Real (1)
Actual Fake (0)        1,862 (TN)            138 (FP)
Actual Real (1)          119 (FN)          1,853 (TP)

True Negatives (Correctly detected Fake): 1,862 articles (93.10% specificity)
True Positives (Correctly identified Real): 1,853 articles (93.97% sensitivity)
False Positive Rate: Only 6.90% of fake stories slipped through.
False Negative Rate: Only 6.03% of real news was flagged as suspicious.

🚀 Predictive System & Inference

The repository includes a ready-to-use prediction function that accepts any headline and body text, formats the input, and returns the classification verdict:

from fake_news import model_pipe, clean_text

def predict_news(title: str, text: str):
    combined = clean_text(f"{title} {text}")
    prediction = model_pipe.predict([combined])[0]
    probabilities = model_pipe.predict_proba([combined])[0]
    
    if prediction == 1:
        print(f"✅ Real News (Confidence: {probabilities[1]:.2%})")
    else:
        print(f"❌ Fake News (Confidence: {probabilities[0]:.2%})")

# Example 1: Authentic Journalism
predict_news(
    title="U.S. lawmakers aim to comply with Iran nuclear deal terms",
    text="WASHINGTON (Reuters) - U.S. lawmakers signaled readiness on Wednesday..."
)
# Output: ✅ Real News (Confidence: 96.40%)

# Example 2: Clickbait / Fabricated Claim
predict_news(
    title="BREAKING: Secret Government Documents Reveal Alien Base Under Pentagon!",
    text="Anonymous insiders claim that underground tunnels were discovered yesterday..."
)
# Output: ❌ Fake News (Confidence: 98.15%)

📁 Repository Structure

fake_news_prediction/
├── WELFake_Dataset.csv   # Benchmark news dataset (72,134 records)
├── fake_news.ipynb       # Interactive notebook (EDA, NLP preprocessing, model, evaluation)
├── requirements.txt      # Python dependencies for exact reproducibility
└── README.md             # Project documentation & benchmark overview

⚙️ Installation & Setup

Prerequisites

Python 3.8+
pip or conda

1. Clone the Repository

git clone https://github.com/devloopcode/fake_news_prediction.git
cd fake_news_prediction

2. Create and Activate a Virtual Environment

Using venv (macOS / Linux):

python3 -m venv venv
source venv/bin/activate

Using venv (Windows PowerShell):

python -m venv venv
.\venv\Scripts\Activate.ps1

Using conda:

conda create -n fakenews-nlp python=3.10 -y
conda activate fakenews-nlp

3. Install Dependencies

pip install --upgrade pip
pip install -r requirements.txt

📖 Usage Guide

Running the Jupyter Notebook

jupyter notebook fake_news.ipynb

Or open the repository in VS Code or Cursor and choose your virtual environment kernel (venv or fakenews-nlp).

🛣️ Roadmap & Enhancements

Baseline text preprocessing and tokenization regex pipeline
Balanced class sampling and EDA distributions
High-dimensional TF-IDF unigram + bigram vectorizer
Scikit-Learn Pipeline integration with Logistic Regression
Interactive predict_news() inference system
Transformer Ensembles: Fine-tuning pretrained language models (RoBERTa, DeBERTa-v3, or DistilBERT) for deep contextual semantic understanding.
Model Explainability: Integrating LIME and SHAP to visually highlight which tokens trigger a "Fake" vs. "Real" verdict.
Web Deployment: Lightweight browser demo using Streamlit or a public REST API with FastAPI.
Source Credibility Scraping: Cross-referencing publisher domains with trusted fact-checking registries (e.g., PolitiFact, Snopes).

🤝 Contributing

Contributions and feature requests are warmly welcomed!

Fork the Project (gh repo fork devloopcode/fake_news_prediction or via GitHub web interface)
Create your Feature Branch (git checkout -b feature/AddRoBERTaModel)
Commit your Changes (git commit -m 'feat: Add RoBERTa text classifier')
Push to the Branch (git push origin feature/AddRoBERTaModel)
Open a Pull Request

👤 Author

Authored with ❤️ by Med-IDBENOUAKRIM
📧 Contact: medidbenouakrim@gmail.com
🐙 GitHub: @devloopcode

_{⭐️ If you found this NLP project helpful, please consider giving it a star on GitHub!}

Navigation Menu

Repository navigation

About

Repository files navigation

📰 Fake News Detection with NLP & Machine Learning

📑 Table of Contents

📌 Executive Summary

🏆 Key Highlights & Performance

📊 Dataset Overview & Profiling

Class Label Convention

Sample Subset & Preprocessing Stats

Linguistic & Length Metrics (Post-Cleaning)

🔬 NLP Pipeline Architecture

🧹 Text Preprocessing & Feature Extraction

1. Robust Regex Text Sanitization

2. TF-IDF Hyperparameters

📈 Model Evaluation & Performance

Classification Report (Holdout Test Set)

Confusion Matrix (Test Set)

🚀 Predictive System & Inference

📁 Repository Structure

⚙️ Installation & Setup

Prerequisites

1. Clone the Repository

2. Create and Activate a Virtual Environment

3. Install Dependencies

📖 Usage Guide

Running the Jupyter Notebook

🛣️ Roadmap & Enhancements

🤝 Contributing

👤 Author

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
README.md		README.md
fake_news.ipynb		fake_news.ipynb
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

📰 Fake News Detection with NLP & Machine Learning

📑 Table of Contents

📌 Executive Summary

🏆 Key Highlights & Performance

📊 Dataset Overview & Profiling

Class Label Convention

Sample Subset & Preprocessing Stats

Linguistic & Length Metrics (Post-Cleaning)

🔬 NLP Pipeline Architecture

🧹 Text Preprocessing & Feature Extraction

1. Robust Regex Text Sanitization

2. TF-IDF Hyperparameters

📈 Model Evaluation & Performance

Classification Report (Holdout Test Set)

Confusion Matrix (Test Set)

🚀 Predictive System & Inference

📁 Repository Structure

⚙️ Installation & Setup

Prerequisites

1. Clone the Repository

2. Create and Activate a Virtual Environment

3. Install Dependencies

📖 Usage Guide

Running the Jupyter Notebook

🛣️ Roadmap & Enhancements

🤝 Contributing

👤 Author

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages