Can I use JMP Pro to do machine learning on spatially referenced (lat, long) data?

[#1]

Hi - I am looking at how best to do machine learning on a set of spatial data (latitude, longitude, variable 1 (numeric), variable 2 (numeric), variable 3 (classified nominal - i.e. can be one of four values).  Can JMP Pro be used for this?  Is there a recommended tutorial or example of this?

My goal is to predict variable 1 based on all the rest of the inputs described above.  Thanks!

I am using JMP Pro 17.

[#2]

The Model Screening tool can test many different modeling techniques at once on your data. @DonMcCormack posted an excellent introduction video here.

[#3]

Thanks Jed.  My main question is whether or not JMP Pro can incorporate the spatial component when modelling the regression.  Will have a look at your link.

[#4]

I watched the intro video and see the robust capabilities of the model screening platform.  However, it is still unclear to me how my x, y coordinates can be part of the available modeled regressions.  The location of my data drives the results (and predicted outcomes) as much as the other variables such as depth to groundwater, etc  (I am predicting contaminant concentration in groundwater as a function of several variables, including distance from the contaminant release.)  So my main question is, can the x, y location be part of the modeled predictions in JMP Pro?

[#5]

Hi @learning_JSL ,

  Yes, JMP Pro can model this. It sounds like you have a single response (Y) variable that you are modeling, and the other variables are all factors (X's) in your model.

  JMP Pro can do this in many ways -- one of which is the Model Screening, which is a neat tool, but it only looks at the default settings for each modeling method, and this is not always the best fit model, so one should keep that in mind when using that platform.

  JMP Pro can model the data using different platforms like Generalized Regression, Boosted Tree, Bootstrap Forest, Neural Net, SVM, K-nearest neighbors, and XGBoost. All of these methods can handle the data you're talking about, and each has their benefits and drawbacks, so you'll really want to model your data using all the different methods and then compare which one does the best at predicting when using a completely different data set that was not used to train and validate the models -- you'll want to save a test data set for this part.

  That being said, you'll also want to split your data into training and validation data sets, Analyze > Predictive Modeling > Make Validation Column, and stratify it on your response column (variable 1, I believe). But do this after you've split off a data table that will be used as your test data to compare the different models against each other.

  For whichever method you use, you will be modeling your response column and you'll be able to save the modeled data as a column formula to the data table. You can then do all different kinds of analyses on the model to determine how good the fit is.

  I'd be happy to run an example for you if you could share your data table. If there is any sensitive information, you could always "anonymize" it: Tables > Anonymize, or even recode a column into categories A,B,C,D if you can't share what the actual entries are.

Hope this helps,

DS

[#6]

Thanks very much DS.  I appreciate the thorough response. My main question is whether JMP Pro can account for data associated with spatial coordinates when it is doing the modeled regressions.  Should I assume that JMP Pro is not capable of incorporating my predictors' locations (each X predictor is associated with a latitude and longitude and the location of the predictors affects how they will effect the response variable Y which also has latitude and longitude)?  So really this is also a spatial regression.   (ArcPro has tools for machine learning that utilize python API, but I was hoping JMP Pro had the capability to do this as I prefer the JMP platofrm over that of ArcGIS.)

[#7 · Accepted solution]

If your concern about the spatial data is the variable grid spacing of geocoordinates, something like a Boosted Neural model might be a good fit, as this would shape the prediction curve to account for this curvature/variability. I've attached a sample workflow that uses JMP's sample data to predict January temperature based on coordinates.

Also, you could likely use the same Python packages directly from JMP, as JMP can run/integrate with Python. Here's a link to a few examples.

[#8]

Excellent.  Will have a look.  Thank you!