What are your favorite data sets for demonstrating Partition Models?

Question

Would love to get your list of go-to's.... thank you in advance!

louv · Answer

I use the Tablet Data as well and also use it to demonstrate the "lock-out" certain solutions capability. I discuss locking out solutions that would require supplier specificity which inevitably would require "sorting" or "cherry picking" lots in the warehouse.

Phil_Kay · Answer

I also use Tablet for partition. I would rarely use anything else.

If I am demonstrating JMP Pro / Predictive Modelling / Machine Learning I would start with Tablet and I might then go to a larger data set. I like the flight delays data. It is not an industrial example but people can relate easily: the first split shows that flight arrival delays are longer after 2.50 pm.

It is good because you can show that partition is quick when you have lots of data. And that you can easily build "better" models with validation and enhanced partition methods. You can also show that graph builder is still very quick with lots of data.

You can download the data from here: http://stat-computing.org/dataexpo/2009/the-data.html

martindemel · Answer

For standard partition my favorite is the tablet production one as it nicely shows in a production case that it can help you to identify what worked best and what didn’t work from historic data. It shows also multi-factor interactions easy to explain to people in the production. Though it is not a full statistical proof that there are no other factors important, my main purpose of showing is:
If you have a production issue and collected some data already, this is a very easy and fast approach to guide the people what to do and not to do based on historic data to gain time for a more thorough analysis! At least you find settings which worked in the past and you can guide your people to produce further , even if not with best optimal settings.

For missing data I use bands data set to show the effect of informative missing. You can use it also for other models

For more advanced situations I use pigment milling dataset (from Phil) which has high mulicolinarities and with standard approaches only temp in the pot is turning out to be significant contradictory to business knowledge. Using bootstrap forest allows for a better selection of potential important factors to inform a new die and finally a better solution than just following the highly linear decreasing temp in pot effect.