Understanding PA results
As you know, ultimately, after working through the entire PA workflow, you will have a single predicted probability of the outcome for each observation in your testing data, which was estimated with your final, selected model/learner.
Based on the performance of the model/learner in your held-out testing data, you will decide whether to deploy your model. We will discuss deployment a little later.
First, there are some valuable questions you may ask to better understand your test results (on new, held-out data) and your final, selected model. Answers to these questions can help with decision-making aimed at improving services, and they can help with explaining your model to stakeholders.
- How are predicted probabilities distributed across and within groups?
- Who has low or high predicted probabilities of the outcome?
- What are the most important predictors in my model?