About WendyCzika

WendyCzika · ‎10-23-2015

Sorry, I was incorrect about the hierarchical clustering above. The Cluster node performs k-means clustering. The optimum number of clusters when using the Automatic setting is first determined making a preliminary cluster pass then k-means is performed with that number of clusters.

WendyCzika · ‎10-21-2015

Download the Files (GitHub) This tip is part of the Learn by Example with SAS® Enterprise Miner™ Templates series where a new data mining topic is introduced and explained with one or more example SAS Enterprise Miner process flow diagrams. When you have a time-dependent outcome that you are trying to model—a failure of some sort or customer churn, for example—you might be interested in predicting when, not if, the event is most likely to occur. In SAS Enterprise Miner, a discrete-time logistic-hazard model is used to perform survival data mining. This approach allows you to model the event likelihood over time, taking into account censored observations, competing risks, time-varying covariates, and left truncation. The hazard function produced by the model can be estimated at a future time interval of interest to answer questions like: What is the probability a customer will churn by a certain date? Which customers are most likely to churn in the next 3 months? What is the expected remaining time for a customer? To get started with survival data mining using SAS Enterprise Miner, download the process flow diagrams (XML files) and the accompanying PDF documentation for the following two examples from the GitHub repository at https://github.com/sassoftware/dm-flow/tree/master/SurvivalAnalysis. Survival: An example that shows the basic use of the Survival node and how different values for the Time Interval property can affect the results SurvivalTVC: An example demonstrating how to include time-varying covariates in your data when modeling with the Survival node. To run these examples, refer to the README file that is part of the GitHub repository at https://github.com/sassoftware/dm-flow. Please note that these examples were tested with SAS Enterprise Miner 13.2. You can find more information about the Survival node in these videos: Introduction to Survival Data Mining New Features in the SAS Enterprise Miner 12.3 Survival Node

WendyCzika · ‎10-21-2015

The Cluster node performs hierarchical clustering using PROC CLUSTER (see the SAS/STAT documentation for more details on that procedure), while the HP Cluster node performs k-means clustering. Here is a helpful tip that can provide you more information on that: https://communities.sas.com/t5/SAS-Communities-Library/Tip-K-means-clustering-in-SAS-comparing-PROC-FASTCLUS-and-PROC/ta-p/221369 Hope that helps! Wendy

WendyCzika · ‎10-09-2015

Another suggestion if you want to keep all the levels is to increase the value for the EM_TRAIN_MAXLEVELS macro variable in the Project Macro Variables window.

WendyCzika · ‎10-08-2015

Download the Files (GitHub) Ever wondered about performing a particular data mining task in SAS Enterprise Miner but weren’t sure how to get started? You can now find templates—example process flow diagrams—focusing on certain applications that you may not be familiar with in a new GitHub repository. An XML file is all you need to easily import a diagram representing a specific task into your SAS Enterprise Miner project. With this imported diagram you can then do the following: examine the results that are created when running the flow with sample data, tweak properties and re-run the flow to see how the results are affected, and run the flow using your own data. A corresponding PDF document is available for each diagram to guide you through the nodes in the flow and the results. Please check out the repository at https://github.com/sassoftware/dm-flow to find templates for these data mining topics: association discovery clustering credit scoring predictive modeling survival data mining text mining time series The README file on the GitHub site contains all the instructions you need to get started. We will feature templates from each of these categories in future articles, so stay tuned!

WendyCzika · ‎09-21-2015

Hi there, It looks like the target has the level interval instead of binary. If you right-click on the Input Data node and select Edit Variables, you can change the level for your target to Binary there, then re-run the StatExplore node. Then in the Results under View>Summary Statistics, you can view the Class Variables: TARGET plot to see the distributions that you want, or the Cell Chi-Squares table. Note that also from the Edit Variables window of the Input Data node, you can select all the variables and click on Explore. This will plot histograms for each individual variable, and these plots are interactive so you can click on the bar for one of the levels in the histogram for your target to see the distribution of the other variables just for that level. Give that a try, and report back how that all goes! Thanks, Wendy

WendyCzika · ‎09-09-2015

If you don't want the formatted values of categorical variables used in the score code, you need to remove the formats from the variables before modeling. Here is information from the reference help: Categories for nominal and ordinal variables are defined by the normalized, formatted values of the variable. If you have not explicitly assigned a format to a variable, the default format for a numeric variable is BEST12., and the default format for a character variable is $w., where w is the length of the variable. The formatted value is normalized by: Removing leading blanks Truncating to 32 characters Changing lowercase letters to uppercase.

WendyCzika · ‎08-27-2015

I think TwoStage will do exactly what you want. Just set both dependent variables as Targets in your Input Data node, and in the TwoStage node, change the Filter property to Non-Events, and Value Model to Tree. You just need to be sure you are using the correct Order (Ascending) for the nominal target so that 0's are treated as the event, then the predicted non-0s will be excluded from the second tree where you are modeling the continuous target.

WendyCzika · ‎08-19-2015

Ray posted a really useful tip about the Control Point node that I think will help: https://communities.sas.com/docs/DOC-6666

WendyCzika · ‎08-14-2015

R_DV is the residual, DV - P_DV. And just a little more about V_DV since that can be confusing. The V_ variables are only calculated for decision trees since for all other models, the V_ and P_ columns would be the same. For decisions trees, V_ represents the proportion of validation obs in the same leaf as the current obs. that have a certain target level for a classification tree, or the mean of the target for the validation obs. in the same leaf as the current obs. for a regression tree.

WendyCzika · ‎08-14-2015

Hi Tony, This message indicates that since the model is predicting almost perfectly (near-0 ASE), that there is likely an input variable that has identical or almost identical values as the target variable. However, it is very strange that you didn't get that message when running with the un-partitioned data. In the Results from the un-partitioned Regression, are you seeing a very small ASE? Thanks, Wendy

WendyCzika · ‎08-14-2015

That's correct about upper-casing the values for the variable measurement levels and roles. But comments and blank lines should be allowed, so not sure why that is causing problems for you. There are some other ways you can define metadata described in this Tip on Defining Global Metadata for SAS® Enterprise Miner™ Projects that might be of interest: https://communities.sas.com/docs/DOC-11122

WendyCzika · ‎08-12-2015

I think it's because the target used in the Tree (the original target) is different than the one used in the Neural Network and Regression nodes (REP_target from the Replacement node). If you give me more detail about how you are using the Replacement node, we can try to come up with a work-around for this. The easiest thing is to connect the DT Model node from the Replacement node instead of the Data Partition node to ensure you are working with the same target, but there is probably a reason you aren't doing that. Wendy

WendyCzika · ‎07-29-2015

Effects are entered into the model based on the most significant p-value from the score chi-square statistic. The process is repeated until none of the remaining effects meet the specified level for entry or until the STOP= value is reached. Then the criterion you are asking about is used to select which step in the selection process is used for the final model. So you should see something like this in the Output window for the DMREG procedure (but note that it won't always be the final step that is selected): Summary of Forward Selection Akaike Effect Number Score Information Step Entered DF In Chi-Square Pr > ChiSq Criterion 1 WOE_PROF 1 1 2273.7985 <.0001 58378.2 2 WOE_STATUS 1 2 1475.0792 <.0001 56842.9 3 WOE_TMJOB1 1 3 842.4397 <.0001 56022.4 4 WOE_TMADD 1 4 274.8227 <.0001 55752.0 The selected model, based on the Akaike information criterion, is the model trained in Step 4. It consists of the following effects: Intercept WOE_STATUS WOE_TMJOB1 WOE_PROF WOE_TMADD

WendyCzika · ‎07-28-2015

This is the value used for the CHOOSE= option in the MODEL statement of PROC DMREG. It is the criterion used to select among the models that are created during the different steps of the model selection; so for example for Forward selection, each of the models created when adding an effect during the forward selection process is evaluated, and the procedure selects the one that is the best in terms of the criterion specified. Hope that helps, Wendy

Online Status	Offline
Date Last Visited	a week ago