📉 Customer Churn Lab — you pick what to predict
Real telecom customer data. Here you choose the output — predict whether a customer churns, or switch the target to their monthly charges, tenure, contract, anything. Then pick the inputs that help and shape the network (max 50 epochs). Binary targets report accuracy; numeric ones report R². The competition is on Churn: set the output to Churn to submit (Score = test MSE × (1 + 0.4% per parameter), up to 3 attempts).
Data: IBM Telco Customer Churn (sample dataset) — 1,500 real customers, seeded 70/30 split. ~28% churned. Some columns (e.g. Total charges) are largely redundant with others — pick what truly helps.
PREDICT (OUTPUT)
INPUTS
Which of the rest help predict it? 3 selected.
Line blue = positive weight, orange = negative; thickness = strength. Watch them shift as it trains.
OUTPUT
• train • test — perfect fit
🧪 Things to try
Change the target
Predict Monthly charges from services, or Tenure from contract & charges. Notice which problems are easy (high R²) and which are hard.
What drives churn?
With Churn as the output, try Contract, tenure, Internet service, Tech support. Month-to-month, fibre, and new customers churn most.
Spot redundancy
Total charges ≈ tenure × monthly. Adding all three rarely beats picking the right one or two — and costs parameters.
Accuracy vs MSE
For Churn, watch both: a model can have decent MSE but poor accuracy if it hedges near 0.5. Push predictions toward 0/1.