# 🌍 Global E-Commerce Customer Purchase Prediction

## Overview

In today’s competitive digital marketplace, understanding customer purchase behavior is critical. E-commerce platforms generate massive amounts of behavioral, transactional, and marketing data — but converting this raw, messy data into actionable predictions remains a challenging task.

In this competition, participants are tasked with predicting whether a global e-commerce customer will make a purchase within the next 30 days.

This is a large-scale, multi-table, intentionally messy dataset designed to simulate real-world industry conditions. Competitors must perform advanced data cleaning, feature engineering, aggregation, and modeling to achieve strong performance.

---

## 🎯 Goal

Build a machine learning model that predicts the probability that a given customer will make a purchase in the next 30 days.

Submissions are evaluated using **ROC-AUC (Receiver Operating Characteristic – Area Under Curve)**.

Your objective is to generate accurate probability predictions between 0 and 1 for each customer in the test set.

---

## 📊 The Challenge

Unlike beginner-friendly tabular competitions, this dataset is intentionally complex and realistic:

- Multi-table relational structure (customers, sessions, transactions, marketing, geo)
- Large-scale data (millions of behavioral records)
- Missing values (including missing-not-at-random)
- Numeric values stored as strings
- Currency-formatted columns
- Inconsistent categorical casing
- High-cardinality features
- Distribution drift between train and test
- Noise and weakly correlated features
- Nonlinear target generation

Success in this competition requires:

- Efficient data processing
- Careful preprocessing and normalization
- Time-aware feature engineering
- Aggregation of behavioral signals
- Robust cross-validation strategy
- Avoidance of data leakage
- Strong modeling techniques (e.g., gradient boosting, ensemble methods)

---

## 🗂 Dataset Structure

The competition data includes multiple relational tables:

- **customers.csv** — customer demographic and profile attributes  
- **sessions.csv** — detailed browsing session logs  
- **transactions.csv** — historical purchase activity  
- **marketing_campaigns.csv** — campaign exposure data  
- **geo_data.csv** — regional socioeconomic attributes  
- **train.csv** — training labels  
- **test.csv** — test set for prediction  

Participants must join and engineer features across tables to build effective models.

---

## 📈 Evaluation

Submissions are scored using:

**ROC-AUC**

Predictions must be probabilities between 0 and 1.

The public leaderboard reflects 30% of the test data.  
Final rankings are determined using the remaining 70% private leaderboard split.

---

## 🚀 Recommended Approach

While many modeling strategies are possible, competitors may consider:

- Aggregating session-level features (recency, frequency, intensity)
- Transaction-derived behavioral signals
- Time-based features and decay modeling
- Handling high-cardinality categorical variables
- Robust handling of missing and noisy data
- Ensemble methods for improved stability

This competition is designed for intermediate to advanced practitioners seeking to strengthen real-world machine learning skills.

---

## 🏆 Who Should Participate?

- Data scientists looking for realistic industry simulation
- Kaggle competitors seeking advanced tabular practice
- Students building production-level ML skills
- Researchers exploring behavioral modeling

If you are ready for a serious, high-complexity tabular challenge — this competition is for you.

Good luck, and may the best model win.