Return Risk Predictor

Check out the live app (deployed on Hugging Face!)

A machine learning-powered Streamlit app that predicts the probability of a product being returned based on customer reviews, delivery metadata, and review ratings.

Overview

Product returns in e-commerce lead to significant losses. This project applies machine learning and natural language processing to predict the return likelihood of a product using customer reviews, metadata, and delivery data.

Built with:

• Python and Streamlit for the UI
• Scikit-learn and XGBoost for machine learning
• TextBlob for sentiment analysis
• Plotly and Seaborn for visualization

Features

Feature	Description
Sentiment Analysis	Uses TextBlob to analyze the tone of customer reviews
Delivery Time Impact	Evaluates how delivery duration affects return chances
Rating Integration	Leverages 1–5 star ratings to gauge satisfaction
Helpfulness Ratio	Measures how helpful other users found the review
Category Encoding	Simulated product category derived from ProductId
Multiple ML Models	Compare predictions using Logistic Regression, Random Forest, and XGBoost
Model Insights	Learn how each model works and what features it relies on

Dataset Used

The project uses the Amazon Product Reviews dataset from Kaggle. Core columns include:

Column	Description
Text	Full customer review
Score	Star rating (1 to 5)
HelpfulnessNumerator	Number of users who found it helpful
HelpfulnessDenominator	Total users who voted
ProductId, UserId	Product and user identifiers
Time	Review timestamp (Unix format)

Additional engineered features:

• delivery_time: Simulated shipping duration
• category_encoded: Encoded first character of ProductId
• review_polarity: Sentiment score using TextBlob
• review_length: Character count of the review
• helpfulness_ratio: Calculated as
  helpfulness_ratio = HelpfulnessNumerator / HelpfulnessDenominator
  (set to 0 when denominator is 0)

Models Compared

Model	Advantages	Use Case
Logistic Regression	Fast and interpretable	Baseline modeling
Random Forest	Handles nonlinear relationships, less overfit	General tabular problems
XGBoost	High accuracy, scalable, feature-aware	Preferred for structured feature data

Performance

Model	Accuracy	AUC Score
Logistic Regression	0.79	0.72
Random Forest	0.84	0.81
XGBoost	0.87	0.89

📽 Demo

Installation

Install dependencies

pip install -r requirements.txt

Run the app

streamlit run app.py

Thank you!