Telco Customer Churn — ANN

Executive Summary

An end-to-end deep-learning pipeline that predicts telecom customer churn from the IBM Telco dataset (7,043 records, 20 features). A compact Keras ANN (64 -> 32 -> 1) is trained with a stratified split and inverse-frequency class weighting to counter the imbalanced churn base rate, then served through a multi-page Streamlit analytics dashboard.

The Problem

Churn datasets are heavily imbalanced — most customers do not churn — so a naive model learns to predict "stay" and misses the customers who actually leave. Retention teams also need per-customer risk, not just an aggregate score, to prioritise outreach.

The Solution

• A feed-forward Keras ANN with a 64 -> 32 -> 1 architecture and dropout for regularisation.
• Inverse-frequency class weighting so the minority (churn) class contributes proportionally to the loss.
• A stratified train/validation/test split that preserves the churn ratio across every fold.
• A Streamlit dashboard for real-time individual risk assessment and batch CSV scoring.

System Architecture

1. Data Layer
   Cleansing and encoding of the 20 raw Telco features (categorical one-hot, numeric scaling).

2. Modeling Layer
   A Keras Sequential ANN (64 -> 32 -> 1, sigmoid output) trained with class weights and early stopping on validation loss.

3. Delivery Layer
   A multi-page Streamlit app with single-record and batch scoring modes that surfaces the main retention drivers behind each prediction.

Engineering Decisions

Why Class Weighting?
Inverse-frequency weighting keeps the imbalanced churn class influential during training without discarding majority-class data through undersampling.

Why an ANN over a Tree Ensemble?
To practise deep-learning workflow design on tabular data — architecture, regularisation, and weighted loss — while keeping the model compact and fast to serve.

Pipeline Details

Dataset: 7,043 records, 20 features (IBM Telco).
Architecture: 64 -> 32 -> 1 feed-forward network.
Split: stratified train/validation/test preserving the churn ratio.
Serving: multi-page Streamlit dashboard (individual + batch scoring).

Outcome

A clean, reproducible churn-prediction pipeline that demonstrates imbalanced-data handling, ANN design, and practical model serving for a retention use case.