Reading Shapley Values for Original Features
- Predictive AI
- H2O Driverless AI
This lesson in the Driverless AI Starter Course demonstrates how to read the Shapley Values for Original Features plot in an MLI report, which ranks dataset features by their contribution to a model's prediction. Using a credit score dataset, it shows how an unexpected top feature, the applicant Name, can signal a data issue rather than a real driver of the outcome, then checks the Datasets tab and finds duplicate applicants under the same name and customer ID. It also covers how Driverless AI excludes ID columns from training by default, a setting that can be changed in Expert Settings.
Key takeaways
- How to read the Shapley Values for Original Features plot to rank which features drive a model's predictions
- Why an unexpected top feature like a customer Name can signal data leakage rather than a genuine predictor
- How to check the Datasets tab for descriptive statistics to spot duplicate applicants under the same name and ID
- The default behavior of excluding ID columns from training, and how to change it in Expert Settings
- What to consider when deciding whether to drop features like Name or Customer ID before rerunning an experiment
Start owning your AI