🕵️ Review Guard

📌 Project Overview

This project was developed by a three-person academic team over 10 weeks on a part-time basis with a $0 budget.

Review Guard is a locally runnable web application that uses Natural Language Processing (NLP) and Machine Learning to classify online reviews as:

• ✅ Genuine
• ❌ Fake

🚀 Features

• Single review prediction
• Batch CSV upload and processing
• Fake/Genuine classification
• Fake probability score output
• Download results as CSV
• Multiple model support (v1, v2, v3)
• Fully offline system

📂 CSV Format

The CSV file must contain the following columns:

text,rating,helpful_vote,verified_purchase

🎯 Project Goals

• Achieve greater than 80% classification accuracy on unseen test data
• Develop a fully offline, locally runnable web application
• Support both single review prediction and batch CSV processing
• Provide fake probability scores for each prediction
• Follow an Agile development methodology with iterative improvements
• Ensure fast response time and user-friendly interface

🛠️ Technologies Used

• Backend: Python, Flask
• Frontend: HTML, CSS, JavaScript
• ML/NLP: Scikit-learn, NLTK
• Data Processing: Pandas, NumPy
• Model Storage: Joblib
• Testing: Pytest
• CI/CD: GitHub Actions

📂 Dataset

Label Definition

• 1 → Fake review
• 0 → Genuine review

The system uses binary classification where 1 represents a fake review and 0 represents a genuine review.

📂 Project Structure

Review-1/
│
├── run.py
├── requirements.txt
├── README.md
├── app/
│   ├── backend/
│   │   ├── __init__.py
│   │   └── app.py
│   ├── frontend/
│   │   ├── templates/
│   │   │   └── index.html
│   │   └── static/
│   │       ├── css/
│   │       └── js/
│   ├── ml/
│   │   ├── __init__.py
│   │   ├── predict.py
│   │   └── training/
│   │       ├── data_processing.py
│   │       ├── feature_engineering.py
│   │       ├── evaluate_model.py
│   │       ├── train_model.py
│   │       └── train_model_v3.py
│   ├── tests/
│   │   ├── test_api_endpoints.py
│   │   ├── test_predict_module.py
│   │   └── test_preprocessing.py
│   ├── artifacts/
│   │   ├── models/
│   │   │   ├── default/
│   │   │   ├── v1/
│   │   │   ├── v2/
│   │   │   └── v3/
│   │   └── reports/
│   ├── data/
│   ├── notebooks/
│   └── src/
└── images/

⚙️ Installation

pip install -r requirements.txt

✅ Run Application

python3 run.py

🏋️ Train Models

Default Training (Recommended)

python app/ml/training/train_model.py \
  --input_csv dataset/amazon_labeled_fake_reviews/final_labeled_fake_reviews.csv \
  --phase1_root app \
  --random_seed 42

Train Model Versions

• v1 (Hybrid): Text + metadata
• v2 (Text-only): Only text features
• v3 (Best Model): Blended model combining v1 + v2

v1

python app/ml/training/train_model.py --include_behavioral --model_version phase1-v1

v2

python app/ml/training/train_model.py --model_version phase1-v2

v3 (Best)

python app/ml/training/train_model_v3.py

Optional (Advanced)

python app/ml/training/train_model.py --enable_xgboost

🧪 Testing

• 106 test cases implemented

• 100% pass rate

• Includes:

  • Unit Testing
  • Integration Testing
  • Security Testing
  • Regression Testing

Run Tests

python3 -m pytest -q app/tests

📊 Model Evaluation

The system is evaluated using:

• Accuracy
• Precision
• Recall
• F1-Score

⚠️ Risk and Validation

To reduce data leakage and improve reliability, the project includes:

• Text normalization and hashing
• Duplicate-safe splitting
• Near-duplicate audit

Remaining Risk

• Paraphrased reviews may still exist in the dataset, creating residual risk

⚙️ CI/CD

• GitHub Actions used
• Automated testing on every push

📊 Project Management (Jira)

The project was managed using Jira with sprint planning and task tracking.

🗂️ Jira Board Overview

📅 Agile Milestones

• Planning and dataset preparation
• Model development and evaluation
• Web application integration
• Testing and final delivery

👥 Team

• Kriti Subedi
• Swapnali Kudale
• Aditi Sharma

📜 License

Developed for academic purposes only.