The Feature Importance list displays the features ordered by how relevant they are to the prediction.

Click a feature to display either a bar or a line chart.

PDPs help us understand how specific values of a feature impacted predictions, and are closely related to SHAP values. Therefore, it can help businesses identify characteristics of problematic or successful behavior in their use case.

How to read PDPs

On all PDPs, the horizontal axis represents the feature's original values, while the vertical axis represents the average effect caused on the prediction.

Let's take a look at a Churn example and consider the hypothetical feature count_distinct_sessions_day (how many times a user logged in each day).

Click Insight and then one of the tags under the feature to display the PDP.

The more sessions a user has each day, the less probable they will churn. If you look at the value of 1 session a day, there is a 36% chance of churning.

By accessing a feature's PDP we gain extra insight into how that feature affects outcomes.

The graph above was typical for a continuous numerical feature. If you had categorical or Boolean features instead (normally in regression models), the PDP would look more like the following:

This is for a demand forecast, and we are displaying the month. We see that there is a higher demand forecast for the last 3 months of the year.

This sort of data enables you to plan special events, etc., to boost demand.

Just like in continuous PDPs, hovering over a point/bar in the PDP will list the exact Effect of Prediction.

You can zoom in on any part of the plot by clicking and dragging the mouse over it.

Did this answer your question?