What business problem does this project address?

Telecom operators lose a large share of subscribers each year with no early-warning signal. The pipeline scores subscribers so retention teams can intervene before cancellation.

How was class imbalance handled?

SMOTE oversampling was applied to the training split only, so validation still reflects the true churn rate and metrics are not inflated.

Which model performed best?

The tuned XGBoost model, reaching roughly 87% accuracy, outperformed Logistic Regression and Random Forest baselines.

Which features mattered most?

Tenure, contract type and monthly charges dominated feature importance, which matches the intuition that short-tenure month-to-month customers on higher bills churn most.

How are the results consumed?

Predictions and driver analysis were packaged into a Power BI retention dashboard so non-technical stakeholders can act on the scores.

Is the notebook public?

Yes, the full pipeline is on GitHub at github.com/Nithin69-k/Telecom-Customer-Churn-Prediction-.

Data Scientist · End-to-end · 2025

Telecom Customer Churn Prediction

An end-to-end ML pipeline that flags at-risk telecom subscribers before they leave.

Pythonscikit-learnXGBoostRandom ForestSMOTEPower BI

View code

Telecom Customer Churn Prediction — An end-to-end ML pipeline that flags at-risk telecom subscribers before they leave.

Overview

an end-to-end telecom churn prediction pipeline with SMOTE balancing, tenure and usage feature engineering, and an XGBoost model reaching about 87% accuracy.

Role: Data Scientist · End-to-end
Year: 2025
Live link: Source on GitHub

Goals

Give operators an early-warning signal before a subscriber churns.
Handle a heavily imbalanced target without collapsing to the majority class.
Identify which factors actually drive churn, not only predict it.
Deliver results in a form a retention team can act on.

Approach & architecture

1
EDA and cleaning
Explored distributions, resolved missing charges and encoded contract and service categories.
2
Class balance
SMOTE oversampling on the training split only, keeping the validation set at the real churn rate.
3
Feature engineering
Derived tenure buckets, usage ratios and contract-type interactions from the raw billing fields.
4
Model benchmarking
Logistic Regression, Random Forest and XGBoost compared on accuracy, recall and ROC-AUC.
5
Explainability
Feature importance ranked the drivers and fed a Power BI retention dashboard.

Machine learning pipeline diagram: raw telecom data, EDA and cleaning, train-test split, SMOTE applied to the training split only, feature engineering, benchmarking of logistic regression, random forest and XGBoost, feature importance and a Power BI dashboard. — End-to-end pipeline — cleaning, SMOTE on train only, benchmarking and dashboarding.

Tech stack

Python

scikit-learn

XGBoost

Random Forest

SMOTE

Power BI

Results

~87%

accuracy from the tuned XGBoost model

actionable churn drivers surfaced: tenure, contract type, monthly charges

model families benchmarked end to end

Telecom Customer Churn Prediction

Overview

Goals

Approach & architecture

EDA and cleaning

Class balance

Feature engineering

Model benchmarking

Explainability

Tech stack

Results

Frequently asked questions

More case studies

J.A.R.V.I.S — NIX

LumoroX Park

LumoroX