I’ve never had success with them. Usually there’s a ton of limitations as well. For anyone who’s used autoML out of the box from one of the big 3 cloud providers, were you able to get it to work well without perfecting the features? I ask it this way because once you do a ton of feature engineering, there’s very little an autoML can do that can increase performance past sklearn out of the box
Working in a startup that is building AutoML platform. Yeah, usually real world datasets are garbage, and shit doesn't work, until you clean it. Once your data is nice, hyperparameter optimization AutoML can help you getting a decent model without much effort. Usually noticeably better than popular baselines, but not as good as a manually tuned model. I was excited initially to join my current company, as they sold me on AutoML. But after some time, and seeing that they don't actually have powerful AutoML in their product, I'm a bit disappointed. Though now have experience of building infrastructure and ML pipelines.
Always establish baseline with automl in a small project, sometime my time is worth more than 5 % improvement. If you ship the model to a billion devices then its different