FakeReview Sentry

FakeReview Sentry is an end-to-end Machine Learning pipeline designed to detect manipulated product reviews on e-commerce platforms. Rather than relying solely on text analysis, it uses a multi-signal ensemble approach that evaluates Natural Language Processing (NLP) metrics alongside reviewer behavioral patterns and metadata.

Architecture & Engineering Focus

This project was built to demonstrate a production-ready ML workflow, from data engineering to model serving.

• Multi-Signal Ensemble: Combines a Naive Bayes NLP model (trained on TF-IDF vectors), a Random Forest behavioral classifier, and an XGBoost model. Predictions are aggregated using a weighted soft-voting mechanism.
• Explainable AI (XAI): Integrated SHAP (SHapley Additive exPlanations) to provide transparency into model decisions, highlighting which specific features (e.g., account age, review velocity, text sentiment) drove the final prediction.
• Optimized Serving: The inference pipeline is decoupled from the training environment. The Flask API and frontend dashboard are containerized via Docker with a minimal dependency footprint (requirements.txt), isolating the heavy training tools (requirements-dev.txt).
• Synthetic Data Engineering: Includes a custom data generation script that models realistic fraud vectors (like burst reviews and bot-network behavior), allowing the ensemble to learn complex interactions that standard Kaggle datasets often lack.

System Components

1. Data Ingestion (data_loader.py): Handles batch CSV loading and data partitioning.
2. Text Preprocessing (text_cleaner.py): NLTK-based pipeline for tokenization, lemmatization, and stop-word removal.
3. Feature Engineering (feature_extractor.py): Computes VADER sentiment, text readability scores, and behavioral velocity metrics.
4. Ensemble Modeling (ensemble.py): Manages model training, evaluation, and serialization.
5. API & Dashboard (app.py, index.html): A REST API serving a clean, responsive web interface for single and batch predictions.

Local Development Guide

1. Environment Setup

git clone <your-repo>
cd FakeReview-Sentry
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements-dev.txt

2. Generate Training Data

Generate the 50,000-row synthetic dataset that includes the necessary behavioral metadata:

python data/generate_sample_dataset.py

3. Model Training

Extract features, train the ensemble, and generate SHAP baseline plots:

python src/models/train_and_evaluate.py --data data/sample_dataset.csv --models-dir models/ --shap

4. Run the Dashboard

python src/api/app.py

Access the dashboard at http://localhost:5000.

5. Run Tests

pytest tests/

Docker Deployment

The application is containerized for easy deployment to cloud providers (e.g., Hugging Face Spaces, Render, AWS ECS).

docker build -t fakereview-sentry .
docker run -p 5000:5000 fakereview-sentry

Model Performance

Evaluated on a stratified 20% holdout test set (10,000 reviews):

• Accuracy: ~89%
• F1 Score: ~87%
• ROC-AUC: ~0.92

Technology Stack

• Core Data Science: Pandas, NumPy, SciPy
• NLP & Features: NLTK, TextBlob, VADER, Textstat
• Machine Learning: Scikit-learn, XGBoost, Imbalanced-learn, SHAP
• Backend & Deployment: Flask, Docker, Pytest