With the advent of Machine Learning, Big data and Artificial Network can models learn over a period of time to take its own decisions?
Asked
Active
Viewed 71 times
1
-
2It is possible that sometime in the future, the fitting of models for only predictive purposes on a completely understood data pipeline will be automated. This would leave the other 90% of the job in human hands. – Matthew Drury Mar 06 '16 at 07:17
-
Not until there is strong AI. At which point, basically everything will be automated. – gung - Reinstate Monica Mar 06 '16 at 10:26
-
@MatthewDrury the key components for that type of automation are already out there at the moment. – Marc Claesen Mar 06 '16 at 12:07
-
See [What do statisticians do that can't be automated?](http://stats.stackexchange.com/q/22572/17230). And "threat" seems to imply a rather pessimistic view of technological progress. – Scortchi - Reinstate Monica Mar 06 '16 at 14:05
1 Answers
1
Well, the greatest part can be automated (and already is) :
- Fitting (cross validating) the model
Many R packages or Python's sk-learn do it quite well.
- Encoding the variables
The factors (per example) can be viewed as dummy variables / integers / mean response of the target and sent to a cross validated model. Thus allowing to choose the best encoding of the features (with respect to the CV score)
- Selecting the best subset of variables
Is really easy to automate as well.
But...
Given the current time needed to train the models (the process presented above is really time consuming), some opinion based choices are necessary
The metric to optimize is a human decision
Feature engineering is hard to automate (and often needs a domain knowledge)
And this second part is usually the one that takes most time...

RUser4512
- 9,226
- 5
- 29
- 59