Me, I find AutoML useful for training an initial baseline model. You know, when you've just managed to get a dataset that's just clean enough to be useful, and you're itching to see if this time you'll need to use something other than LightGBM. It's always LightGBM though. Always. It's that good.
Anyway, I don't really use AutoML apart from this initial phase. It's slow enough to make it impractical for regular retrains, and when you need a model that's constantly updated, well, slow is out and fast is in.
What I do instead is reverse engineer the AutoML model, take whatever I can from there and adapt whatever I can't take. Kinda like Picasso, I imagine 🙂.
If you're curious about the process, I've written a short(er) blog post on creating your own models based on Azure AutoML-trained ones. You'll find it here. It's focused on time series, but the underlying principles should apply to regression and classification models, too.
Hope you like it. And I appreciate any and all feedback you may have - just reply directly to this email, or head over to Twitter and let me know.
Thanks for reading,