Create Testing Set (Data Mining Wizard)

Use the Create Testing Set page to specify how much of the data is to be used for training, and how much is to be reserved for use as a test set. Separating data into a training and testing set when you create a mining structure makes it much easier to assess the accuracy of mining models that you create later.

You can specify the amount of testing data as a percentage, or you can specify a number to limit the number of cases used for testing. If you specify both a percentage and a maximum number of cases to use for testing, both settings are used, and the testing data set contains the smaller number of cases. By default, 30 percent of data is used for testing, 70 percent for training, and there is no maximum number of test cases.

By default, Analysis Services generates a numeric seed that is used to start partitioning. This seed is based on the name of the mining structure. If you want to ensure that the partition stays the same even if the name of the mining structure is changed, you can specify a value for the seed, by setting the HoldoutSeed property of the mining structure. If you change the holdout seed, you must reprocess the structure.

If you later want to change the amount of testing or training data, you can modify the HoldoutMaxCases and HoldoutMaxPercent properties on the data mining structure by using the Properties window. However, after you make the change you must reprocess the mining structure and all associated mining models. The following limitations also apply:

• Partitioning of a data mining structure is supported only when the data mining structure is stored in SQL Server 2008. Earlier versions of SQL Server Analysis Services do not support caching of partition information for mining structures.

• You cannot partition a mining structure if the mining structure contains a Key Time column, which is required for time series mining models.

• You cannot partition data if you are trying to predict a value that is stored in a nested table.

For More Information:Validating Data Mining Models (Analysis Services - Data Mining), Creating a New Mining Structure, Basic Data Mining Tutorial