[marcinzm]

>In general, if your DS's can't explain the model then something has gone horribly wrong.

With modern ML models data scientists can generally explain why the model should work but not exactly how it gives any specific prediction. Explaining how a model with a trillion parameters comes up with an answer is not a trivial task.

[disgruntledphd2]

> modern ML models

As an aside, this is a hilarious phrasing. What are we going to call these methods in a decade? You'd probably be better off phrasing it as deep neural networks.

Anyway, this is not really true. There are methods (for lots see: https://christophm.github.io/interpretable-ml-book/) and the DeepDream paper came out in 2015, so it's possible.

It's computationally expensive, and a lot of people don't see the value, but my argument is that if you want to use the model in the real world, and have non-technical stakeholders then you'll need to do this, and in general, I've found it to be the best way to actually improve a model.

And to be fair, if you just need to see how the predictions vary as a function of the inputs, you can again hold all but one constant and run a bunch of different values through the model.

Again, this can all be done, but I think it's more a question of will than capability (and hence the context of my original commment).

[marcinzm]

>You'd probably be better off phrasing it as deep neural networks.

I phrased it the way I did for a reason. Large Random Forrest models are also not easily interpretable. Even large logistic regression models with feature interactions and feature hashing aren't easily interpretable. It's not a question of the model technique used but rather the amount of parameters and how many feature interactions are modeled.

>And to be fair, if you just need to see how the predictions vary as a function of the inputs, you can again hold all but one constant and run a bunch of different values through the model.

This only provides a partial view of the feature impact due to non-linear interactions.