HN Explorer

Ask HN: Automated machine learning products live up to the promise?

2 points by eyeball ↗ HN
My company is looking at some “AI” products like datarobot and h2o driverless. The sales pitch for these is that they automate away a lot of the hard stuff of data science and will enable a less technical “business analyst” to effectively create and deploy machine learning models.

Curious if anyone can share experiences with this. Has it really worked out that way?

The team members that would use such a tool, should we buy one, are fairly data savvy, can do the data prep work/etl; and have good domain knowledge for the problems we would try to model. wondering if they could be successful at the model building work with a tool like this despite lacking real undestanding / know how to do such work “manually”.

2 comments

[ 3.4 ms ] story [ 12.7 ms ] thread
grizzles
IDK about those products but Google's AutoML stuff is probably the best of breed here. A few research groups have published AutoML-like stuff on github too.

Like anything it depends on your use case. Sometimes even a tiny bit of feature engineering can have a huge impact on model fidelity.

eyeball
The google one looked like it was focused only on a few use cases (language / vision).

I've played around with some of the open source stuff (tpot, autosklearn, h20's automl).

The commercial products seem to do more automated feature engineering.

H2o's automl is pretty easy to use with their graphical "flow" notebook interface. Was able to build a reasonable model with no coding aside from prepping the data going in.