Partial Dependence & ICE Plot Explorer: Understanding Model…

What Partial Dependence Plots Are

Partial dependence plots (PDPs) are graphical representations that show the marginal effect of a feature on the predicted outcome of a machine learning model. By averaging the model's predictions over all other features, PDPs help visualize how changes in one feature affect the model output.

The plot essentially averages out the influence of all other variables to isolate and display the relationship between the selected feature and the prediction.

Why ICE Curves Fan Out

Individual conditional expectation (ICE) curves fan out because they represent the model's predictions for each individual instance in the dataset, rather than an average. This allows us to see how different instances respond differently to changes in a feature.

The fanning out of ICE curves indicates that the relationship between a feature and the prediction can vary significantly across different data points.

live demo · related simulation● LIVE

How Interaction Strength Affects PDPs

When interaction strength is high, it means that the effect of one feature on the model's output depends strongly on the value of another feature. This can cause the partial dependence plot to show more complex patterns and non-linear relationships.

In contrast, when there are no interactions or low interaction strength, the PDP will typically show a simpler linear relationship between the selected feature and the prediction.

The Role of Instance Count

Increasing the instance count in a partial dependence plot can provide more accurate averages by including more data points. However, it also increases computational complexity and may smooth out individual variations.

A higher number of instances can help capture the true distribution of responses but might obscure local patterns if not enough smoothing is applied.

Frequently asked questions

What does a fanning out ICE curve indicate?

A fanning out ICE curve indicates that different data points in the dataset respond differently to changes in the feature being analyzed, highlighting variability and non-uniformity in the model's behavior.

How can I use PDPs and ICE curves together effectively?

By comparing PDPs with ICE curves, you can identify whether the model's predictions are consistent across different instances or if there is significant variability. This helps in understanding both the average behavior and individual differences in the model.

Why might I want to toggle ICE curves on/off?

Toggling ICE curves allows you to see the range of responses from individual data points, which can help identify outliers or instances where the model behaves unexpectedly. This is particularly useful for detecting anomalies in your dataset.

How does interaction strength affect my PDP?

Higher interaction strength causes the partial dependence plot to show more complex patterns and non-linear relationships, reflecting how the effect of one feature on the model's output depends on the value of another feature.

Try it live

Everything above runs in your browser — open Partial Dependence & ICE Plot Explorer and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.

▶ Open Partial Dependence & ICE Plot Explorer simulation

Partial Dependence & ICE Plot Explorer: Understanding Model Behavior