Intuition

1. Let’s select a subset of observations similar to observation of interest, also known as neighbours.

2. Now, we can compare the distribution of residuals of the neighbours and the entire training dataset except for neighbours by creating a local-fidelity plots or running a statistical tests.

\[\begin{equation} r_i = y_i - f(x_i) = y_i - \hat y_i \end{equation}\]

3. We can also check how the “small differences” in explanatory variables (present in the neighbours) can influence on the predictions across the range of each variable by comparing the CP profiles of each observation by creating local-stability plots.