MLE is a fairly new title that, as best I can tell, exists primarily in those few places that have a mature enough workflow to have people who can actually dedicate their time to the ML part and have other roles take care of the rest.
Everywhere else, there is only DS, and it involves everything.
To answer your first question though, the training and testing of these models is fun because it feels like a puzzle game: did all my understanding and preparation of the data (and the business) pay off and the model does its job as expected? Is there something I’m missing? What’s the simplest model + configuration I can use that produces acceptable results and what does that say about the problem space? Can I combine models in some way to get the results? Is nothing working because it’s an ultimately fruitless exercise and our hypothesis is wrong? Or is there something we’re missing that is in turn the reason the model is missing something? Etc etc.
Then as the output you get something that ingests some data and then makes a decision with it! That’s cool to me.
I get where you're coming from. I guess just the problem domain I'm in, and my experience level, I tend to get what I expect from a model, and if I don't I'm more like, "wtf?" which isn't anywhere as fun of a way to do that part of the process.
Also, I know what is possible and impossible before I start writing code (if you don't count EDA code). There are exceptions, like it should be possible but it turns out the data is bad, but it didn't look bad from the EDA. Thankfully I've never had that. I always perform a Feasibility Assessment before anything else.
Not to imply what you're doing is somehow incorrect. Problems can vary quite a bit and I recognize that. For example, there have been times where I've had to mine to see if anything is there, doing ML over it to validate a hypothesis then using that information to create a new hypothesis, rinse and repeat. That's scary, because I could turn up nothing. I haven't done a lot of mining I admit though. Usually my problems are much more obvious from the get go, or much more research intensive.
One time I did three months of reading papers on arxiv.org just to figure out if something was feasible and how to best do it. Though that was definitely not a standard problem.
Everywhere else, there is only DS, and it involves everything.
To answer your first question though, the training and testing of these models is fun because it feels like a puzzle game: did all my understanding and preparation of the data (and the business) pay off and the model does its job as expected? Is there something I’m missing? What’s the simplest model + configuration I can use that produces acceptable results and what does that say about the problem space? Can I combine models in some way to get the results? Is nothing working because it’s an ultimately fruitless exercise and our hypothesis is wrong? Or is there something we’re missing that is in turn the reason the model is missing something? Etc etc.
Then as the output you get something that ingests some data and then makes a decision with it! That’s cool to me.