AutoTrain MLInput features
19 service, account, billing, and demographic fields
Machine learning use case
Customer churn prediction
Build and compare customer churn classifiers using contract, service, tenure, and billing features.
Experiment design
Setup used for the reference run
This fixed setup makes the comparison understandable and repeatable. It is an example configuration, not an automatic recommendation for every dataset.
DatasetIBM Telco Customer Churn
TargetChurn (Yes/No)
Validation80/20 holdout
Random state42
Preprocessing
Coerce TotalCharges to numeric, impute missing values, standardize numeric columns, one-hot encode categories.
Build sequence
Recreate the experiment in AutoTrain
- Inspect the dataset.Confirm field types, missing values, class balance, identifiers, and whether each feature exists at prediction time.
- Save preprocessing.Apply the documented transformations without fitting them on the holdout rows.
- Train comparable runs.Keep the dataset version and split constant while changing only the algorithm or intended hyperparameters.
- Review errors.Use task-appropriate metrics and inspect where the model fails before creating a testing tab.
Fixed reference run
Reference benchmark
These values were computed with scikit-learn using the setup above. They show what one reproducible baseline looks like; they are not copied from the cited paper and are not a production guarantee.
Accuracy by model
Higher is betterF1 for churn class: 0.604, 0.535, and 0.589.
Reading the result
What the benchmark means
The two strongest accuracy values are tied, but Logistic Regression has the highest churn-class F1 in this fixed run. That makes it the clearer baseline here; a final choice should still compare recall, calibration, and retention cost.
Benchmark vs. your experiment: the chart above is a static reference. Results shown inside your dashboard are generated from your own dataset, parameters, validation split, and saved model.
Visual context for the customer churn prediction workflow. Model selection should be based on validation metrics and product constraints.
Limitations and risk
This is a fictional telco sample. Class imbalance makes accuracy insufficient by itself; review churn-class recall, precision, calibration, and the cost of retention actions.
Before deployment
Repeat validation on recent, representative data; review subgroup errors; define a fallback; and monitor input and outcome drift.
Human decisions
Use the prediction as evidence, not as the sole decision maker, where an error could affect education, health, access, pricing, or customer treatment.
Journal foundation
Verbeke et al. (2012), New insights into churn prediction in the telecommunication sector, European Journal of Operational Research.
The journal reference provides methodological or domain context. AutoTrain's benchmark above is a separate implementation using the stated holdout configuration.
Open journal article