Fake_Review_Detection

🛡️ Fake Review Detection System using BERT & Explainable AI (XAI)

A production-grade NLP pipeline that detects fake/spam reviews on online platforms using a fine-tuned BERT (bert-base-uncased) transformer model and interprets predictions using Explainable AI (SHAP & LIME) to highlight word-level feature contributions.

---

Problem Statement

Online marketplaces and review platforms (Amazon, Yelp, Google Reviews) are frequently targeted by automated bot networks and fake reviewers posting deceptive feedback. Traditional keyword filtering and rule-based systems fail to handle complex semantic patterns, subtle hyperbole, and contextual spam.

This project addresses fake review detection using Deep Transformer Architecture (BERT) combined with Explainable AI (XAI), providing not only state-of-the-art detection accuracy but also human-understandable explanations for why a review was flagged as fake or genuine.

---

Objectives

• Binary Classification: Classify customer reviews into Genuine (0) or Fake/Spam (1).
• Baseline Modeling: Build a classic TF-IDF + Logistic Regression benchmark for comparison.
• Transformer Fine-Tuning: Fine-tune pre-trained bert-base-uncased on labeled review data using PyTorch.
• Explainable AI (XAI): Integrate SHAP and LIME to generate word-level feature importance visualizations for individual predictions.
• Comprehensive Evaluation: Compare models across Accuracy, Precision, Recall, F1-Score, and AUC-ROC curves.

---

Tech Stack

Category	Library / Tool	Purpose
Core Language	Python 3.10+	Primary development language
Deep Learning	PyTorch & HuggingFace Transformers	BERT model fine-tuning & sequence classification
Baseline ML	Scikit-Learn	TF-IDF vectorization, Logistic Regression, evaluation metrics
Explainable AI (XAI)	SHAP & LIME	Local and global word importance explanations
Data & EDA	Pandas, NumPy, WordCloud	Preprocessing, feature extraction, N-grams
Visualization	Matplotlib, Seaborn	Confusion matrix plots, ROC curves, SHAP summary plots

---

Directory Structure

Fake_Review_Detection/
├── fake_review_detection.ipynb   ← Main Jupyter notebook (EDA, Models, XAI, Inference)
├── README.md                     ← Project documentation & results summary
├── requirements.txt              ← Dependencies
└── data/
    └── fake_reviews.csv          ← Benchmark dataset (text, label, rating, verification)

---

Dataset Overview

• Source & Public Dataset Links:
  • Fake Reviews Dataset (Kaggle)
  • Amazon Product Reviews Dataset (Kaggle)
  • Yelp Review Dataset (Kaggle)
• Local Benchmark File: data/fake_reviews.csv

The dataset contains 1,200 labeled reviews with balanced target classes:

• review_id: Unique identifier for each review.
• text: Raw review text.
• label: 0 for Genuine Review, 1 for Fake/Spam Review.
• rating: Customer rating (1 to 5 stars).
• verified_purchase: Binary indicator (1 = Verified Buyer, 0 = Unverified).

Key EDA Insights

1. Length Distribution: Genuine reviews tend to be longer with specific descriptions of product features and minor drawbacks, whereas fake reviews show a bimodal length distribution (either short repetitive spam or exaggerated reviews).
2. Word Clouds & Vocabulary:
   • Genuine Reviews: Words like battery, quality, design, setup, days, performance, price, minor, decent.
   • Fake Reviews: Words like best, buy now, 10/10, coupon, super, discount, click link, 5 stars, cheap, scam.

---

Model Architecture & Methodology

1. Baseline Model (TF-IDF + Logistic Regression)

• Text tokenized using TfidfVectorizer (unigrams + bigrams, top 5,000 features).
• Logistic Regression trained with L2 regularization.

2. Advanced Model (Fine-Tuned BERT)

• Base checkpoint: bert-base-uncased.
• Architecture: 12-layer Transformer encoder with sequence classification head.
• Tokenization: BertTokenizer with max_length=128, padding, and truncation.
• Training Setup: AdamW optimizer ($lr=2\times 10^{-5}$), linear warmup scheduler, trained for 3 epochs with gradient clipping.

---

Explainable AI (XAI) Integration

SHAP (SHapley Additive exPlanations)

• Quantifies exact game-theoretic Shapley values for each word's contribution toward shifting the model's output probability from baseline to predicted outcome.
• Positive SHAP values (pushing toward Fake): "buy now", "discount", "coupon", "10/10", "fast ship".
• Negative SHAP values (pushing toward Genuine): "battery", "weeks", "decent", "however", "packaging".

LIME (Local Interpretable Model-agnostic Explanations)

• Fits local surrogate models around specific review samples.
• Highlights word importance weights dynamically in real-time inference.

---

Experimental Results & Performance Comparison

Model	Accuracy	Precision	Recall	F1-Score	AUC-ROC
TF-IDF + Logistic Regression	98.33%	98.35%	98.33%	98.33%	0.9995
Fine-Tuned BERT (bert-base-uncased)	99.58%	99.59%	99.58%	99.58%	1.0000

---

How to Run

1. Clone the repository and navigate to the project directory

git clone https://github.com/Niketkumardheeryan/ML-CaPsule.git
cd ML-CaPsule/Fake_Review_Detection

2. Install dependencies

pip install -r requirements.txt

3. Launch Jupyter Notebook

jupyter notebook fake_review_detection.ipynb

4. Real-Time Inference Example

You can test custom review strings directly in the notebook using the built-in predict_and_explain(text) pipeline:

review = "BEST PRODUCT EVER WOW AMAZING DEAL BUY NOW 10/10 FIVE STARS!"
result = predict_and_explain(review)
# Output: Label: Fake/Spam | Confidence: 99.87% | Key Drivers: 'buy now', '10/10', 'amazing'

---

Why This Project Stands Out

• Production Relevant: Solves a major real-world security and trust problem in e-commerce.
• Interpretability First: Moves beyond black-box classification by incorporating Explainable AI (XAI).
• Comprehensive Benchmarking: Compares traditional NLP baselines against state-of-the-art Transformer models.
• Interview-Ready: Well-documented modular codebase suitable for machine learning portfolios.

---

License

Distributed under the MIT License. See LICENSE in the root repository for details.