Machine learning excels because it's not interpretable. Not in spite of it. Interpretability is a constraint.Thanks for reading Mindful Modeler! Subscribe for free to receive new posts and support my work. Every constraint shrinks the hypothesis space of models, making it less likely to find the best-performing one.
For non-tabular datasets, what are the available model-agnostic post hoc IML methods that are comparable to methods like ALE for tabular datasets? Is this the subject of Chapter 10 of your IML book (https://christophm.github.io/interpretable-ml-book/neural-networks.html) or do you have more to say on this topic?
For tabular datasets, I would think that model-agnostic post hoc methods like ALE resolve this conflict: go all out for performance, then interpret the best-performing model. From that perspective, I understand you to be saying that we should focus almost exclusively on performance while model-building and then use suitable interpretation methods afterwards, rather than being constrained by limiting ourselves to only using "interpretable" models while we are still in development phase. Is that the basic gist of your article (from the tabular dataset perspective)?
Hi Christoph, I can see where you are coming from. My concern is when interpreting the model is key. For example being able to clearly articulate why someone didn't get the loan to a regulator.
The Case for Uninterpretable Machine Learning
One might argue (per Breiman) that statisticians never had the wheel to begin with! ML has always been a practical affair.
For non-tabular datasets, what are the available model-agnostic post hoc IML methods that are comparable to methods like ALE for tabular datasets? Is this the subject of Chapter 10 of your IML book (https://christophm.github.io/interpretable-ml-book/neural-networks.html) or do you have more to say on this topic?
For tabular datasets, I would think that model-agnostic post hoc methods like ALE resolve this conflict: go all out for performance, then interpret the best-performing model. From that perspective, I understand you to be saying that we should focus almost exclusively on performance while model-building and then use suitable interpretation methods afterwards, rather than being constrained by limiting ourselves to only using "interpretable" models while we are still in development phase. Is that the basic gist of your article (from the tabular dataset perspective)?
Hi Christoph, I can see where you are coming from. My concern is when interpreting the model is key. For example being able to clearly articulate why someone didn't get the loan to a regulator.