Review Fraud Detection System

Overview

This project implements a machine learning system that classifies product reviews as either fraudulent or non-fraudulent. The goal is to help consumers identify authentic reviews that provide reliable information about products they're considering purchasing.

The system uses a two-stage approach:

1. Sentiment Classification: Determines whether a review is positive or negative
2. Fraud Detection: Identifies reviews that are likely to be fraudulent (computer-generated)

Key Features

• BERT-based deep learning models fine-tuned for review classification
• Custom data loaders and preprocessing pipelines
• Dual model approach (with/without review titles)
• CLI tool for testing fraud detection on custom reviews
• Comprehensive evaluation metrics and visualization

Project Structure

Fraud Classification/
├──App/
│   ├── config/              # Holds config.yaml, which contains hyperparams, etc...
│   ├── model/               # Holds the model that is being used to classify fraud
│   ├── cli.py               # The way you will interact with the model
│   └── model_def.py  
├── BERT_model/
│   │   ├── Fraud Classification          # Training for Fraud Model
│   │   ├── Sentiment_Analysis            # Training for sentiment classification Model
├── requirements.txt         # Project dependencies
└── README.md                # This file

Installation

Requirements

• Python 3.8+
• CUDA 12.8+ (optional, for GPU acceleration)

Dependencies

The project requires the following libraries:

torch>=1.7.0  
transformers>=4.5.0  
pandas>=1.0.0  
numpy>=1.19.0  
scikit-learn>=0.24.0  
matplotlib>=3.3.0  
seaborn>=0.11.0  
pyyaml>=5.4.0  
huggingface_hub[hf_xet]>=0.23.0

Standard Installation

pip install -r requirements.txt

Manual Installation

pip install torch>=1.7.0 \
            transformers>=4.5.0 \
            pandas>=1.0.0 \
            numpy>=1.19.0 \
            scikit-learn>=0.24.0 \
            matplotlib>=3.3.0 \
            seaborn>=0.11.0 \
            pyyaml>=5.4.0 \
            'huggingface_hub[hf_xet]>=0.23.0'

GPU Support

For GPU acceleration, install the CUDA-compatible version of PyTorch:

# First, install CUDA 12.8 from NVIDIA's website
# Then install the CUDA-compatible PyTorch version

Usage

Dataset Preparation

1. Sentiment Analysis Datasets:

   • Download the Amazon Review Polarity CSV dataset
   • Place it in the datasets/ directory corresponding to amazon_review_sentiment

2. Fraud Classification Dataset:

   • A sample dataset is provided in datasets/fraud_data.csv (which is only needed by fraud classifier training)
   • The below link is for if you want to run any of the files in processed_datasets/amazon_review_dataset_2018
     • For full reproduction, download the 2018 Amazon Review Dataset

Training Models

Sentiment Classification

# Train the model with review titles
cd BERT_model/Sentiment_Analysis
python train.py

# Train the model without review titles
python train_no_title.py

Fraud Classification

cd BERT_model/Fraud_Classification
python train.py

After training, select the best model checkpoint (typically the one with the highest validation accuracy) and move it to the model/ directory.

Testing Fraud Detection

cd App
python cli.py

This launches an interactive CLI where you can enter a review text and rating to test if it's classified as fraudulent.

Results

The current implementation achieves:

• 94.3% accuracy on sentiment classification with review titles
• 92.1% accuracy on sentiment classification without titles
• 100% validation accuracy on fraud detection (epoch 3 model)

Future Improvements

• Increase the training dataset size for fraud classification
• Incorporate product category stratification
• Deploy as a web service with API
• Experiment with alternative model architectures
• Detect human-written fraudulent reviews (not just computer-generated)

Note

This project is an initial proof of concept with satisfactory results. Due to computational resource constraints, models were trained on a limited dataset. The approach demonstrates viability, but would benefit from additional data and computing resources.