Related Posts
I got a remote L6 SWE offer at Google but the RSUs are too low for that level ($625k over 4 years), could not negotiate it higher. I’ve accepted the offer but trying to be hopeful that I could make it more equitable in the future if I perform well. Are stock refreshers based on the initial RSU grant or solely based on performance? Google
Pro
I don’t trust cheese that doesn’t melt
Is it too late to buy Sandbox land now?
Additional Posts in Data & Analytics Consultants
What’s everyone’s take on Alteryx?
Has anyone else begun to resent data science?
New to Fishbowl?
Download the Fishbowl app to
unlock all discussions on Fishbowl.
unlock all discussions on Fishbowl.
I have yet to see a true no code ML solution ever be used by someone without ML experience
Bowl Leader
I've used several. I suggest you experiment, but don't stake your career in it.
Used H2O.ai to hyper-parameter tune boosted trees and basic NN's on my pre-rolled data. Still ran through the R API and needed some ML understanding, so I didn't see a no-code use case. As results go, I pretty consistently got equal to slightly better metrics with feature engineering on more robust/simpler models. Can't hurt to keep around as a boilerplate script to benchmark with, but that's about it.
A niche vendor with an actual no code solution did better. Pull up a customer list, click a few buttons, wait a while, and soon you have a stopgap model. The metrics weren't good and you could forget about it being clean, but still it usually gave some lift in practice--good enough for the business while we made a better model over however long. Plus, we negotiated some data access, which was great for BI (generally too noisy for augmentation in models, but interesting nevertheless). Tread lightly, and make it obvious to pilot teams that this is a radically different off the cuff model just to get things moving, not your actual prototype.
One thing to beware of is that most niche vendors will want to run everything black box on their servers. This means you have to expose some of your data, which is often a problem. How to work this out is situational. I'd ask vendors for references on how other customers solved similar issues, and talk to legal about whether the proposed solution is ok.
That was incredibly insightful. So the summary is it can do an ok job at driving short term results in the business domain. Although customized scripts will usually outperform them.
It was an oversight on my part regarding data security. But it is noted now.
Thank you very much again!
Are the tools just not good enough yet?
This is a big topic - I am surprised there are not more replies!
The various AutoML approaches (H20 and autosklearn) are fantastic packages to get the data scientist up and running quickly when handed a data set. I encourage my team to use [various flavors of] AutoML as a baseline for any modeling work that we do.
It is very cheap to set up and can take you pretty far with a few lines of code. Do I want my team to spend their time chasing down parameters, or fiddling with this vs that model? Absolutely not. I want my team of scientists doing science - I want them chasing down the big questions that unlock better model performance, adoption, understanding, [and] or pivot/shift their efforts quickly to the next thing from the business.
I am a fan of "if you are going to use the tool, we should know the underlying magic".
Auto-sklearn: https://arxiv.org/abs/2007.04074