Hi Marc,

Yes, as I understand it, the party package does use conditional inference trees—that is its main raison d'etre—and they will be used regardless of whether you enable the “Use conditional variable importance” option in MGET. That option is not utilized during model fitting, only after the model is fitted and variable importance is being estimated. Specifically, it is the value passed by MGET to the party package’s varimp() function for the conditional parameter. The relevant documentation is:

Function varimp can be used to compute variable importance measures similar to those computed by [the function] importance. Besides the standard version, a conditional version is available, that adjusts for correlations between predictor variables.

If conditional = TRUE, the importance of each variable is computed by permuting within a grid defined by the covariates that are associated (with 1 - p-value greater than threshold) to the variable of interest. The resulting variable importance score is conditional in the sense of beta coefficients in varimp regression models, but represents the effect of a variable in both main effects and interactions. See Strobl et al. (2008) for details.

Note, however, that all random forest results are subject to random variation. Thus, before interpreting the importance ranking, check whether the same ranking is achieved with a different random seed – or otherwise increase the number of trees ntree in ctree_control.

I attached some relevant papers.

I have had mixed success with conditional variable importance. It seems that the implementation for it is very memory intensive. Because MGET runs inside ArcGIS’s traditional 32-bit Windows applications (ArcMap, ArcCatalog), it can only utilize the 32-bit version of R. In practice, this limits MGET to about 1 to 2 GB of memory, regardless of how much is available from the operating system itself. Memory-hungry code can quickly blow through this. It seems surprising that 900 records x 9 predictors would do that, although if there is some code that would, it is the conditional importance code.

If it is a borderline situation you might be able to get it to succeed by starting a fresh copy of ArcGIS and then running the tool without doing anything else first. Also, you could try running it from ArcCatalog instead of ArcMap; that might result in more memory being available to R that would otherwise be taken up by ArcMap for the map itself, etc.

At the end of the day, if you believe the conditional variable importance is the way to go, your best bet is probably to fit the model yourself in a 64-bit version of R. That would allow more memory to be accessed by the algorithm. It is not a guaranteed solution but might work.

Best,

Jason

From: [mailto:] On Behalf Of Marc Nelson
Sent: Tuesday, March 21, 2017 8:47 AM
To:
Subject: [mget-help] Question about Conditional Interference Trees - 'Fit Random Forest Model' tool

Hi Jason / team,

/Marc