Choose the Column to Use for Testing a Mining Model
Applies To: SQL Server 2016
Before you can measure the accuracy of a mining model, you must decide which outcome it is that you want to assess. Most data mining models require that you choose at least one column to use as the predictable attribute when you create the model. Therefore, when you test the accuracy of the model, you generally must select that attribute to test.
The following list describes some additional considerations for choosing the predictable attribute to use in testing:
Some types of data mining models can predict multiple attributes—such as neural networks, which can explore the relationships between many attributes.
Other types of mining models—such as clustering models—do not necessarily have a predictable attribute. Clustering models cannot be tested unless they have a predictable attribute.
To create a scatter plot or measure the accuracy of a regression model requires that you choose a continuous predictable attribute as the outcome. In that case, you cannot specify a target value. If you are creating anything other than a scatter plot, the underlying mining structure column must also have a content type of Discrete or Discretized.
If you choose a discrete attribute as the predictable outcome, you can also specify a target value, or you can leave the Predict Value field blank. If you include a Predict Value, the chart will measure only the model’s effectiveness at predicting the target value. If you do not specify a target outcome, the model is measured for its accuracy in predicting all outcomes.
If you want to include multiple models and compare them in a single accuracy chart, all models must use the same predictable column.
When you create a cross-validation report, Analysis Services will automatically analyze all models that have the same predictable attribute.
When the option, Synchronize Prediction columns and Values, is selected, Analysis Services automatically chooses predictable columns that have the same names and matching data types. If your columns do not meet these criteria, you can turn off this option and manually choose a predictable column. You might need to do this if you are testing the model with an external data set that has different columns than the model. However, if you choose a column with the wrong the type of data you will either get an error or bad results.
Double-click the mining structure to open it in Data Mining Designer.
Select the Mining Accuracy Chart tab.
Select the Input Selection tab.
On the Input Selection tab, under Predictable Column Name, select a predictable column for each model that you include in the chart.
The mining model columns that are available in the Predictable Column Name box are only those with the usage type set to Predict or Predict Only.
If you want to determine the lift for a model, you must select a specific outcome value to measure, for by choosing from the Predict Value list.