Rapid Predictive Modeler is really useful. In just a few clicks, you can create a model that serves as a great starting point for any data mining project. You can then tweak this model into a better predictive model that suits your needs. Plus, the report it generates helps you understand the main drivers of your RPM model.
In this article, you’ll learn how to interpret the results of your RPM model. If you’re looking for a guide on how to import your RPM model into SAS Enterprise or just want to learn more about RPM, check out 's article .
For this example, imagine that you need to create a model using RPM to identify customers who have a high probability of defaulting on their credit payments. The data set you will use is the German Credit table found in the library sampsio. This sample data set contains inputs to model a binary target called good_bad, which flags all customers that defaulted on their credit payments.
Rapid Predictive Modeler Results
Once you have run your rapid predictive model, open the PDF report that it generates. You will find several summary tables and graphs for your model.
The tables and graphs in your report help you understand the model that was selected. This means that different settings of RPM on the same data set will give you different tables and graphs as your report depends on the final model that RPM selected, not on your data set.
Selected Variable Importance and Cross Tabulations Scorecard
Two of the most useful summary results are the Selected Variable Importance and the Cross Tabulations Scorecard. The Selected Variable Importance graphic shows you which variables contribute the most to your chosen model whereas the Cross Tabulations Scorecard helps you determine which values within those significant variables have the most effect. These two reports used together provide you with the whole picture about which variables, and which values within that variable drive your RPM model.
How does Rapid Predictive Modeler Calculate Selected Variable Importance?
The variable importance in this plot is calculated using a decision tree algorithm to explain the predicted outcome variable of your RPM model. By default, the predicted outcome I_<target> (i_good_bad in this example) is used as target and all variables flagged as significant in the selected model as inputs. When you specify a decision matrix, the decision outcome chosen by the model (d_good_bad in this example) is used.
This plot is particularly useful to explain black box models like neural networks or support vector machines. RPM’s tree-based variable importance will not necessarily match the variable importance of methods that already calculate variable importance like regression, random forest, gradient boosting, etc.
Selected Variable Importance Interpretation
Take a look at the Selected Variable Importance of this example. It indicates that there is one variable that is very significant in your model. The variable checking (years with checking account) is a very strong driver in your model to predict payment default. This information is very useful to us; we can now use it to dictate adequate delinquency policies and create new strategies for cross-sale and customer retention.
Checking isn't the only important variable here. Several other variables contribute to your predictions and you’ll want to take a closer look at them, especially history, duration, amount and savings.
Figure 1. RPM Selected Variable Importance graph
How does Rapid Predictive Modeler create the Cross Tabulations Scorecard?
The variables are first binned by the same decision tree algorithm used to determine the Selected Variable Importance above. Next the binned variables are used as inputs to a regression model using the predicted outcome as target. By default the regression uses i_<target> as the dependent variable, or d_<target> if you specified a decisions matrix. In our example, the regression uses the binned inputs of the german credit data set to explain the variable i_good_bad calculated by the selected RPM model.
The scorecard points are calculated through a scaling that starts by identifying the lowest parameter estimate of the regression within a binned variable. This value is assigned a score of 0 and will be used as a reference value. The scorecard points for all other binned levels are scaled based on the difference between the parameter estimate of that bin and the parameter estimate of the reference level. The scorecard points values range from 0 to 1000 and increase as the difference between the parameter estimate and the reference parameter estimate increases. The more similar the parameter estimates across all binned levels within a variable, the closer they will be to 0. On the other extreme, if there is a binned level that explains most of a variable, and has a very high parameter estimate compared to the other binned levels, the associated score will be closer to 1000.
Cross Tabulation Scorecard Interpretation
When interpreting the scorecard for Rapid Predictive Modeling results, you’ll have to make a clear distinction between a scorecard generated through SAS Credit Scoring for SAS Enterprise Miner and one from RPM. These two scorecards are not the same. The scorecard produced by the Scorecard node is a true scorecard in the sense that points are generated in terms of certain scaling properties, and are comparable across variables.
Since the scorecard from Rapid Predictive Modeler is based on a different scaling algorithm, the points from one variable are not comparable to the points of another variable. However, Rapid Predictive Modeler Cross Tabulation Scorecard gives you a clear notion of the relationship of certain values of your inputs with the event you are modeling.
Once you have identified significant variables from the Selected Variable Importance graphic above, you’re now able to dive deeper and determine what specific values, or range of values, are controlling the significance of the variable.
In our example, you learned that the variable checking (years with checking account) was the most important variable in your RPM model. You can now look to the Cross Tabulation Scorecard and use the scorecard points to notice that the higher the number of checking accounts, the lower the chances of a customer being bad. You can also notice this relationship by comparing the 43.54% bad rate for bin 1 to the 20.57% bad rate of bin 3. Notice as well that there are twice as many customers in bin 3, compared to bin 1, which is an indicator of a good quality portfolio.
Figure 2. Partial screen of RPM Cross Tabulation Scorecard
This example helps you understand better the algorithms and the logic behind two of the most useful results generated by the SAS Rapid Predictive Modeler. You should also be more familiar with the advantages and limitations of these results, and how to use them together to gain better insights about your selected model.
If you find this tip helpful, have any questions, or simply want to share your thoughts, please comment below.
Registration is open! SAS is returning to Vegas for an AI and analytics experience like no other! Whether you're an executive, manager, end user or SAS partner, SAS Innovate is designed for everyone on your team. Register for just $495 by 12/31/2023.
If you are interested in speaking, there is still time to submit a session idea. More details are posted on the website.
Data Literacy is for all, even absolute beginners. Jump on board with this free e-learning and boost your career prospects.