Review Fraud Detection System

This project implements a multi-layer review fraud detection system with:

• Textual layer (50% weight)
• Behavioral layer (30% weight)
• Temporal layer (20% weight)

It integrates customer, product, and review datasets, computes a Final Fraud Risk Score (0–100), aggregates scores at the product level, and exposes an interactive Streamlit dashboard.

---

1. Project Structure

The expected structure is:

fraud_detection_project/
│
├── data/
│   ├── customers.csv                      # or customer_preprocessed_only_200_rows.csv
│   ├── products.csv                       # or product_preprocessed_only_200_rows.csv
│   └── reviews.xlsx                       # or textual_temporal_200_rows.csv.xlsx
│
├── models/
│   └── fraud_model.py
│
├── utils/
│   ├── textual_layer.py
│   ├── behavioral_layer.py
│   ├── temporal_layer.py
│
├── app.py
├── setup_env.py
├── requirements.txt
└── README.md

Note: If you prefer to keep your original filenames, just copy them into the data folder. The code will automatically try both the generic names (customers.csv, products.csv, reviews.xlsx) and the original filenames (customer_preprocessed_only_200_rows.csv, product_preprocessed_only_200_rows.csv, textual_temporal_200_rows.csv.xlsx).

---

2. Environment Setup (venv + Auto Install)

From the fraud_detection_project folder:

2.1 Create virtual environment venv

python -m venv venv

On Windows (PowerShell):

.\venv\Scripts\Activate

On macOS / Linux:

source venv/bin/activate

2.2 Install dependencies from requirements.txt

You can either install manually:

pip install -r requirements.txt

or let the automation script do everything for you (recommended):

python setup_env.py

The setup_env.py script will:

• Create the venv virtual environment (if it doesn’t exist)
• Install all dependencies from requirements.txt inside venv
• Download NLTK data:
  • punkt
  • stopwords
• Download spaCy model:
  • en_core_web_sm

---

3. Expected Data Files

Place your files in the data folder. The loader will automatically try multiple naming options:

• Customers
  • data/customers.csv
  • data/customer_preprocessed_only_200_rows.csv
• Products
  • data/products.csv
  • data/product_preprocessed_only_200_rows.csv
• Reviews
  • data/reviews.xlsx
  • data/textual_temporal_200_rows.csv.xlsx

Required logical columns (names can vary slightly; see code for fallbacks/aliases):

• Customer_ID
• Product_ID
• Review_ID (or equivalent)
• Review_Text
• Optional: Text_Fraud_Probability
• Behavioral:
  • Account_Age
  • Review_Frequency
  • Refund_Ratio
  • Verified_Purchase_Ratio
  • Average_Rating_By_User
• Temporal:
  • Reviews_Per_Day
  • Burst_Flag
  • Review_Date

---

4. Running the Streamlit Dashboard

After the environment is ready and data files are in data/:

streamlit run app.py

The dashboard provides:

• Dropdown selection by Review_ID
• Display of:
  • Review text
  • Textual score
  • Behavioral score
  • Temporal score
  • Final weighted fraud risk score
  • Risk level (Low / Moderate / High)
• Product-level metrics:
  • Average fraud score per product
  • Suspicious review ratio per product
  • Product authenticity score
• Basic charts:
  • Distribution of fraud scores
  • Per-product fraud metrics

---

5. Reproducibility Notes

• Random operations (e.g. train/test splits, model initialization) use a fixed random seed for reproducibility.
• All key hyperparameters and column names are defined in one place in the code and can be adjusted easily.