5.3 Exploration of global structure with Ceteris Paribus profiles

Profiles presented in previous sections are also useful for exploration of the global structure of a model. This may be useful to

  • assess the average model response for a selected variable,
  • identify interaction between variables,
  • assess the variable importance.

Once neighbors are identified, one can plot profiles for every neighbor. In the Figure 5.3 the blue curve shows Ceteris Paribus profiles for a variable of interest, while all other profiles are for the neighbors.

Profiles are parallel here, what suggests that model is additive in respect to the surface variable. Profiles are not far from each other, what suggests that the observation is stable.

Figure 5.5: a1


Figure 5.6: a1


Figure 5.7: a1


Figure 5.8: a1

a1 How to do this in R?

First, one need to identify neighbors for selected points of interest. This can be done with the select_neighbours() function. By default it is using the gower distance, but one may change it with the distance argument.

Here we select 10 closest points.