SelfExplainML/PiML-Toolbox

Support residual by level or bin of feature?

jphall663 opened this issue · 1 comments

Would be a nice enhancement to support residual plots and analysis by feature and by level or bin of feature. This just allows for spotting of specific bugs, like in the image below.

mlha_0807

Prototype code here: https://nbviewer.org/github/jphall663/interpretable_machine_learning_with_python/blob/master/debugging_resid_analysis_redux.ipynb, cell 18.

For categorical features, residuals can be plotted across top-k levels, for numeric features they will have to binned before plotting.

Thank you for this valuable suggestion.

Currently, the residual plot is based on full samples. We will consider adding the option to visualize only a subset of samples to be customized by users.