🕵️‍♂️ Mission Impostible – Group 17

✨ Who We Are

Welcome to Mission Impostible, a team of emerging data scientists from around the globe who are boldly navigating the world of analytics — while quietly battling an all-too-familiar foe: Imposter Syndrome.

The name? A cheeky nod to how we feel sometimes. But while we may be new to the game, we’re not backing down. We’re collaborating, learning, and growing together—one dataset at a time—and we’re on a mission to tackle real-world problems with data.

---

📚 Table of Contents

• 🛒 Problem Background: Returns in E-commerce
• 📦 Returns Prediction: Graph-Based vs Tabular E-Commerce Modeling
• 🧠 Modeling Approaches: Why Two Methods?
  • A. Tabular Modeling (TheLook Dataset)
  • B. Graph Modeling (ASOS Dataset)
• 📁 Dataset Documentation
  • A. ASOS GraphReturns Dataset . B. TheLook E-Commerce Dataset
• 📊 Summary Table
• 🛍️ E-commerce Product Return Analysis – Data Preparation
  • 📦 ASOS GraphReturns – Data Preparation
  • 📦 TheLook E-commerce – Data Preparation
• 📁 Folder Structure
• 📈 Data Exploration & Insights
  • ASOS GraphReturns
  • TheLook E-commerce
• 🤖 Predictive Modeling & Results
  • ASOS (Graph-Based)
  • TheLook (Tabular)
• 💡 Key Takeaways
• 👨‍👩‍👧‍👦 Team of Imposters
• 📜 License

---

🛒 Problem Background: Returns in E-commerce

Product returns are a major challenge for online retailers, costing companies billions annually—not just in lost sales, but also in reverse logistics, inventory disruption, fraud, and environmental waste. Managing returns impacts profitability and customer satisfaction, requiring robust data-driven solutions.

---

📦 Returns Prediction: Graph-Based vs Tabular E-Commerce Modeling

This project compares two advanced approaches to predicting product returns in e-commerce:

• Graph-based modeling (ASOS Returns Prediction)
• Tabular data modeling (TheLook E-Commerce)

Our goal: Understand and predict product returns using customer, product, and transaction features—empowering smarter business decisions.

---

1. 🧠 Modeling Approaches: Why Two Methods?

Predicting returns is crucial for e-commerce logistics and customer experience. We explored:

📦 A. Tabular Modeling (TheLook Dataset)

• Each row = one purchase (features: product, customer, price, shipping, etc.)
• Models: Logistic Regression, XGBoost
• Strengths: Simple, interpretable, fits business rules
• Limitations: Treats transactions independently

🧩 B. Graph Modeling (ASOS Dataset)

• Nodes: Customers & products
• Edges: Purchases (labeled as returned/not)
• Model: Graph Neural Networks (GNNs)
• Strengths: Captures shared patterns, e.g., return-prone users or products
• Limitations: No time data, only includes customers with returns

---

2. 📁 Dataset Documentation

A. ASOS GraphReturns Dataset

• Source: OSF
• Data: Edge list with return labels, anonymized features
• Notes: No timestamps, only customers with returns
• How to get: Download from OSF

B. TheLook E-Commerce Dataset

• Source: Kaggle
• Files: order_items.csv, products.csv, users.csv, distribution_centers.csv
• Notes: Some missing timestamps, synthetic PII
• How to get: Download from Kaggle

---

📊 Summary Table

Model Type	Good For	Not Good For
Tabular	Easy to use & interpret	Finding relationships
Graph	Complex pattern detection	Needs dense data, harder to explain

Both methods provide valuable insights—use them together for best results.

---

🛍️ E-commerce Product Return Analysis – Data Preparation

We engineered modeling-ready datasets for both approaches:

📦 ASOS GraphReturns – Data Preparation

• Input: .p files (event, customer, product nodes)
• Process: Merge, clean, rename columns
• Output: asos_merged_training.csv
• How to run: Use 01_data_preparation.ipynb in Google Colab

📦 TheLook E-commerce – Data Preparation

• Input: Four CSVs (orders, products, users, centers)
• Process: Clean, engineer features (return flag, discount %, basket size, tenure, shipping latency)
• Output: thelook_returns_features.csv
• How to run: Use theLookdata_preparation.ipynb in Jupyter/VS Code

---

📁 Folder Structure

.
├── 0_domain_study
├── 1_datasets
├── 2_data_preparation
├── 3_data_exploration
├── 4_data_analysis
├── 5_communication_strategy
├── 6_final_presentation
├── collaboration
├── notes
├── .github
├── .vscode
├── LICENSE
├── README.md
└── ... (other configuration files)

---

📈 Data Exploration & Insights

ASOS GraphReturns

• Return rates: Most purchases are not returned; most customers return only once or twice.
• Key drivers: Product type, gender, country, age.
• Imbalance: Most records are "not returned".
• Visuals: Return frequencies, rates by demographic, product, geography.

TheLook E-commerce

• Return share: ~10% of items returned.
• Key drivers: Season, product category, country, distribution center.
• Numeric features: Weak linear correlation with returns—non-linear models needed.
• Visuals: Feature distributions, return rates by group.

---

🤖 Predictive Modeling & Results

ASOS (Graph-Based)

• Models: Logistic Regression, Random Forest, GNNs
• Top features: Customer/product return history, product type, country
• Accuracy: ~75% (Random Forest)
• Limitations: No return reasons, only historical patterns

TheLook (Tabular)

• Models: Logistic Regression (baseline), XGBoost (advanced)
• Best ROC-AUC: 0.655 (XGBoost)
• Top features: Product category, discount %, basket size, tenure
• Insights: High error in certain categories/geographies; numeric features alone not enough

---

💡 Key Takeaways

• Graph and tabular models each reveal unique patterns.
• Product category, customer history, and geography are strong predictors.
• Combining both approaches can improve return prediction and business strategy.

---

👥 Team of Imposters:

• Shadi Shadabshoar — LinkedIn
• Pyae Linn — LinkedIn
• Derek Karungani

---

📜 License

MIT License

---

© 2025 | Mission Impostible – E-commerce Product Return Prediction Project