topic Re: HP GLM not working with Model Comparison Node in SAS Data Science

HP GLM not working with Model Comparison Node

Mike90 — Tue, 07 Nov 2017 20:22:15 GMT

If I connect two HP GLM nodes to a Model Comparison node, the node fails, with the error message "ERROR: Must use one target variable modeled by a predecessor node." The HP GLM nodes have exactly 1 target variable, as indicated in the Variable Summary in the Results.

If I connect one HP GLM node and one decision tree node to the Model Comparison node, it only uses the decision tree.

The partition is 40/30/30.

SAS Enterprise Miner 14.2

This document shows an HP GLM node directly in between a Data Partition node and a Model Comparison node. This indicates a special code node is not needed. "What’s New in SAS Enterprise Miner™ 13.1." https://support.sas.com/resources/papers/proceedings14/SAS311-2014.pdf

Re: HP GLM not working with Model Comparison Node

WendyCzika — Wed, 08 Nov 2017 12:04:19 GMT

When you specify a discrete distribution for the Interval Target Probability Distribution property, model assessment is not performed and model comparison with the Model Comparison node is unavailable.The Fit Statistics table from the HPGENSELECT procedure shown in the Output window can help you compare between 2 HP GLM runs.

Re: HP GLM not working with Model Comparison Node

Mike90 — Wed, 08 Nov 2017 17:39:32 GMT

The target is of type Interval, and is discrete, as it consists of integers from 1 to around 1100 (days), and is an exponential distribution. I don't need discrete predictions.

I ran the HP GLM node with the default settings. Changing the target probability distribution from Poisson to Tweedie results in fit statistics, and the results appear in the Model Comparison node.

The help pages mention what you said about target probability distributions that are discrete, but they fail to say which ones are discrete. I did a few Google searches, and I'm not finding that information. I'll keep looking. I could make models with the other 6 types to determine this, but I'd like more detailed information.

Re: HP GLM not working with Model Comparison Node

WendyCzika — Wed, 08 Nov 2017 17:43:41 GMT

Poisson, Negative binomial, and the ZI versions of those are the discrete ones.

Re: HP GLM not working with Model Comparison Node

Mike90 — Wed, 08 Nov 2017 18:54:33 GMT

Thanks. I've found some documentation on this now.

I don't get why necessary data isn't passed on for model comparison, but at leas I now know that Model Comparison won't work with the discrete target probability options.

Re: HP GLM not working with Model Comparison Node

WendyCzika — Wed, 08 Nov 2017 19:10:58 GMT

Actually I just tried this work-around that worked: put a Model Import node in between HP GLM and Model Comparison, then I think it will include HP GLM for discrete distributions in your Model Comparison.

Re: HP GLM not working with Model Comparison Node

Mike90 — Thu, 09 Nov 2017 16:48:19 GMT

This works.

(The Model Comparison node uses the name of the import node, which has a default of MdlImp,)