16.3 Interpretation trade-off

In prior chapters we have been using model-specific as they are closer tied to the model performance, but you would be facing the next limitations:

• Not all models have a method. For example, the stacked model we trained during last chapter.
• Comparing feature importance across model classes is difficult since you are comparing different measurements.

There isn’t a Best Approach

The only way to full trust our model interpretations is to apply multiple approaches including model specific and model agnostic results.