Diagnostics for classification models

This chapter discusses different diagnostics for classification models in a model agnostic fashion. It first shows how missclassifications by a given model can be explored, and how we can distinguish if the missclassification is a reasonable result of class overlap, or a sign of an inadequate model. It then describes how we can visualise boundaries between predicted classes, and how this can be useful when comparing different classification models for the same data. A final section is then introducing local explainers, demonstrated based on Shapley values and how they can be used to further understand the way a model is operating in the neighbourhood of individual observations.