churn-intel-ai

Predict. Segment. Retain. ML system built on 5,630 customers achieving 99.4% AUC-ROC. Quantifies ₹7.8M revenue at risk, simulates retention campaigns with upto 6,032% ROI & delivers insights via an interactive real-time Plotly Dash dashboard.

🧠 churn-intel-ai

AI-Driven Customer Churn Prediction & Revenue Retention Analytics

📌 Project Overview

churn-intel-ai is an end-to-end machine learning system designed to predict customer churn in an e-commerce environment — before it happens. The system goes beyond standard classification by integrating Customer Lifetime Value (CLV) modelling, revenue-at-risk quantification, retention ROI simulation, and most importantly — per-customer personalised action intelligence that tells a business exactly who to contact, why, and what to offer them.

Built on a dataset of 5,630 customers across 20 behavioral and transactional features, the system achieves a near-perfect AUC-ROC of 99.4% using XGBoost, identifies over ₹7.8 Million in revenue at risk, and auto-generates retention strategies for the top 50 highest-risk customers based on their individual behavioral parameters.

⚙️ Tech Stack

Layer	Tools
🐍 Language	Python 3.11+
📊 Data Processing	Pandas, NumPy
📉 Visualization	Matplotlib, Seaborn, Plotly
🤖 Machine Learning	Scikit-Learn, XGBoost, Imbalanced-Learn
💹 Financial Modelling	Custom CLV & Revenue-at-Risk Engine
🎯 CTA Engine	Rule-based Per-Customer Action Generator
📡 Dashboard	Plotly Dash
🗃️ Data Format	Excel (.xlsx), CSV
🔧 Environment	VS Code, Jupyter Notebooks, virtualenv

🗂️ Repository Structure

churn-intel-ai/
│
├── 📁 data/
│   ├── raw/                          # Original dataset (E_Commerce_Dataset.xlsx)
│   └── processed/                    # Cleaned, engineered & scored CSVs
│
├── 📓 notebooks/
│   ├── 01_EDA.ipynb                  # Exploratory Data Analysis
│   ├── 02_preprocessing.ipynb        # Feature Engineering & Data Prep
│   ├── 03_modeling.ipynb             # ML Model Training & Evaluation
│   ├── 04_clv_analysis.ipynb         # CLV + Revenue at Risk Calculation
│   └── 05_financial_simulation.ipynb # Retention ROI Simulation
│
├── 📁 src/
│   ├── preprocess.py
│   ├── model.py
│   ├── clv.py
│   └── retention_strategy.py
│
├── 📊 dashboard/
│   └── app.py                        # Interactive Plotly Dash Dashboard
│
├── 📁 outputs/
│   ├── models/                       # Saved .pkl model files
│   └── reports/                      # Generated charts & CSV reports
│
└── 📄 requirements.txt

🔬 System Architecture

Raw Data (Excel)
      │
      ▼
┌─────────────────────┐
│  EDA & Visualization │  ──► Churn patterns, correlations, risk indicators
└─────────────────────┘
      │
      ▼
┌──────────────────────────┐
│  Preprocessing & Feature  │  ──► Null imputation, encoding,
│  Engineering              │       EngagementScore, LoyaltyTier,
└──────────────────────────┘       ComplaintRisk, HighValue flag
      │
      ▼
┌────────────────────────────────────┐
│  ML Pipeline                        │
│  ├── Logistic Regression (AUC 86.7%)│
│  ├── Random Forest    (AUC 97.9%)   │
│  └── XGBoost ✅       (AUC 99.4%)   │
└────────────────────────────────────┘
      │
      ▼
┌──────────────────────────┐
│  Churn Probability Scores │  ──► Risk Segments: High / Medium / Low
└──────────────────────────┘
      │
      ▼
┌──────────────────────────┐
│  CLV + Revenue at Risk    │  ──► Per customer financial impact
└──────────────────────────┘
      │
      ▼
┌──────────────────────────┐
│  Retention ROI Simulation │  ──► 4 campaign scenarios modelled
└──────────────────────────┘
      │
      ▼
┌────────────────────────────────────┐
│  Per-Customer CTA Intelligence      │  ──► 8 risk rules → personalised
│  Engine (NEW)                       │       action per customer
└────────────────────────────────────┘
      │
      ▼
┌──────────────────────────┐
│  Plotly Dash Dashboard    │  ──► Live interactive business intelligence
└──────────────────────────┘

📊 Key Results & Findings

🤖 Model Performance

Model	Accuracy	Precision	Recall	F1 Score	AUC-ROC
Logistic Regression	80.8%	82.3%	81.9%	58.6%	86.7%
Random Forest	94.7%	82.5%	84.2%	83.3%	97.9%
XGBoost ✅	96.5%	93.6%	85.8%	89.6%	99.4%

XGBoost selected as final production model. Confusion matrix shows only 27 missed churners out of 1,126 test samples.

🔍 Top Churn Drivers (Feature Importance)

Rank	Feature	Type
🥇 1	LoyaltyTier_Enc	Engineered
🥈 2	Tenure	Original
🥉 3	Complain	Original
4	HighValue	Engineered
5	ComplaintRisk	Engineered
6	NumberOfDeviceRegistered	Original
7	PreferedOrderCat_Enc	Encoded

3 out of top 5 features were custom engineered — validating the feature engineering strategy.

📉 Critical Business Insights

⚠️  New customers (0-3 months)     →  36.2% churn rate  [CRITICAL ZONE]
✅  Customers with 24+ months       →  0.0%  churn rate  [FULLY LOYAL]
🔴  Customers who complained        →  31.7% churn rate
🟢  Customers with no complaints    →  10.9% churn rate
📱  COD payment users               →  28.8% churn rate  [HIGHEST RISK]
💳  Credit Card users               →  12.9% churn rate  [LOWEST RISK]
👤  Single marital status           →  26.7% churn rate
📍  Tier 3 city customers           →  21.4% churn rate

💰 Financial Analysis Summary

Metric	Value
💼 Total CLV Portfolio	₹84.1 Million
🚨 Total Revenue at Risk	₹11.3 Million
🔴 High Risk Segment Revenue at Risk	₹7,853,629
📉 12-Month Revenue Loss (no action)	~₹10 Million
🛡️ Revenue Saved (Premium Campaign)	₹3,017,094
📈 Best Campaign ROI	6,032%
⏰ Cost of Inaction Per Month	₹850,000

🧪 Retention Scenario Simulation

Scenario	Customers Targeted	Revenue Saved	Net Gain	ROI
❌ No Intervention	0	₹0	₹0	0%
📧 Basic Campaign (20%)	All High Risk	₹3,201,654	₹3,041,454	1,898%
🎯 Targeted Campaign (40%)	High Value + High Risk	₹2,290,698	₹2,241,498	4,556%
👑 Premium Campaign (60%)	High Value + High Risk	₹3,017,094	₹2,967,894	6,032%

🎯 Per-Customer CTA Intelligence Engine

This is the most business-critical feature of the project. Rather than just identifying at-risk customers, the system auto-generates a personalised retention action for each of the top 50 highest-risk customers based on their individual behavioral risk parameters.

8 Risk Rules Powering the CTA Engine

Rule	Trigger Condition	Action Generated
🆕 New Customer	Tenure ≤ 3 months	Assign onboarding manager, welcome kit within 48 hours
📢 Filed Complaint	Complain = 1	Priority support escalation, service recovery coupon
😞 Low Satisfaction	SatisfactionScore ≤ 2	Satisfaction recovery email, free delivery for 3 orders
🔴 Critical Risk	ChurnProbability ≥ 0.85	Immediate phone call, personalised 3-month retention deal
⏳ No Recent Activity	DaySinceLastOrder = 0	Re-engagement push notification within 12 hours
💎 High Value at Risk	BusinessSegment = High Value High Risk	VIP account manager, loyalty tier upgrade
⚡ Sudden Drop Signal	DaySinceLastOrder ≤ 2 + ChurnProb ≥ 0.70	Investigate last order experience proactively
🚚 High Delivery Distance	WarehouseToHome ≥ 25	Free express delivery offer for next 5 orders

What the CTA Table Shows Per Customer

⚠️ Risk Reasons — Every specific flag that triggered their inclusion in the list
✅ Recommended Action — Exact step the retention team should take
Red highlighted cells — The precise column value that caused the flag
Sortable & Filterable — Filter by complaint = 1 to get only complainers instantly

📊 Dashboard Features

The Plotly Dash dashboard (dashboard/app.py) provides a complete business intelligence interface:

📌 Section 1 — KPI Overview

Total customers, high-risk count, revenue at risk, avg CLV, overall churn rate

⚡ Section 2 — CTA Recommendation Cards

6 company-level action cards derived directly from model insights
Each card maps to a specific data finding with a concrete business action

📉 Section 3 — Analytics Visualizations

Churn probability distribution with decision threshold line
Risk segment donut chart
CLV vs Churn Probability scatter (color-coded by risk)
Revenue at Risk by segment bar chart
Dynamic categorical churn rate explorer
Financial simulation dual-axis chart (Net Gain + ROI)

🎯 Section 4 — Top 50 Priority Retention Table

Top 50 highest revenue-at-risk customers
Red cell highlighting on exact risk-triggering columns
Sortable and filterable live table

🧠 Section 5 — Per-Customer CTA Intelligence

5 summary KPIs (callbacks needed, complaints, new customers, high value at risk, total stake)
Full per-customer risk reasons + recommended action table
Tooltip support for full action text
Color-coded: orange for risk reasons, green for actions

🚀 Getting Started

1️⃣ Clone the Repository

git clone https://github.com/divyansh1920/churn-intel-ai.git
cd churn-intel-ai

2️⃣ Create Virtual Environment

python -m venv .venv
.venv\Scripts\activate        # Windows
source .venv/bin/activate     # Mac/Linux

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Add Dataset

Place E_Commerce_Dataset.xlsx inside:

data/raw/E_Commerce_Dataset.xlsx

5️⃣ Run Notebooks in Order

notebooks/01_EDA.ipynb
notebooks/02_preprocessing.ipynb
notebooks/03_modeling.ipynb
notebooks/04_clv_analysis.ipynb
notebooks/05_financial_simulation.ipynb

6️⃣ Launch Dashboard

cd dashboard
python app.py
# Open → http://127.0.0.1:8050

📦 Requirements

pandas
numpy
matplotlib
seaborn
scikit-learn
xgboost
imbalanced-learn
plotly
dash
jupyter
ipykernel
openpyxl

Install all at once:

pip install -r requirements.txt

📁 Output Files Generated

File	Location	Description
`df_scored.csv`	data/processed/	All customers with churn probability + risk segment
`df_clv.csv`	data/processed/	CLV + Revenue at Risk per customer
`xgb_model.pkl`	outputs/models/	Trained XGBoost model
`scaler.pkl`	outputs/models/	Fitted StandardScaler
`imputer.pkl`	outputs/models/	Fitted SimpleImputer
`label_encoders.pkl`	outputs/models/	Fitted LabelEncoders
`priority_retention_list.csv`	outputs/reports/	Top 20 high-risk customers
`financial_simulation_results.csv`	outputs/reports/	All 4 scenario results
`model_comparison.csv`	outputs/reports/	Model metrics comparison
Charts 01–10 (.png)	outputs/reports/	All visualization outputs

🧠 Feature Engineering — Custom Features Created

Feature	Formula / Logic	Purpose
`EngagementScore`	Weighted combo of HourSpendOnApp, OrderCount, CouponUsed	Overall activity score
`RevenuePotential`	CashbackAmount × OrderCount	Revenue proxy per customer
`ComplaintRisk`	Complain == 1 AND SatisfactionScore ≤ 2	High-risk behavioral flag
`LoyaltyTier`	Tenure binned → New / Growing / Established / Loyal	Tenure-based segment
`HighValue`	CashbackAmount > 75th percentile	Binary high-value flag

📌 Dataset Information

Property	Detail
Source	Kaggle — E-Commerce Customer Churn Dataset
Sheet	`E Comm`
Rows	5,630 customers
Features	20 original + 5 engineered
Target	`Churn` (0 = Retained, 1 = Churned)
Churn Rate	16.8%
Class Imbalance Handling	SMOTE (Synthetic Minority Oversampling)

👤 Author

Divyansh Jain 🔗 LinkedIn 🐙 GitHub

📄 License

This project is licensed under the MIT License — feel free to use, modify, and distribute with attribution.

Built with 🧠 machine learning, 💰 financial analytics, 🎯 per-customer intelligence, and ⚡ real business thinking.

Name		Name	Last commit message	Last commit date
dashboard		dashboard
data		data
notebooks		notebooks
outputs		outputs
src		src
README.md		README.md
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation