As Raman mentioned in the Data Miner newsletter, the November CTP of SQL Server 2008 is now available.
It includes many really cool new features in Analysis Services. Among them: Holdout support, Model filtering, DMX Column aliasing, Drillthrough enhancements, Cross validation and practically two new forecasting algorithms under the Microsoft_Time_Series umbrella.
I intend to present all of them briefly, and I start today with the Holdout support.
Most of the data mining tasks require a validation step, performed right after modeling. This validation step
consist (typically) in evaluating model’s performance against data that was not seen during training (test data).
The test data ideally has the same statistical properties as the training data (data seen by the mining model during training). An easy way to achieve statistical similarity between the training and test set is to use random sampling. This method is not guaranteed to give correct results (statistical similar populations) but, assuming that the random sampling mechanism is independent of the data, it will work in most common scenarios.
SQL Server Integration Services has a Random Sample transform, which extracts a random sample (with a certain percentage) using a mechanism independent of the actual data being samples. This is why we strongly recommended using Integration Services to generate test/training partitions for SQL Server Data Mining.
However, there are a few problems:
- Integration Services will have to save at least one of the sample sets in a relational table (or some form of output destination)
- This sampling method can only be applied to data coming from a relational source (or, in general, a source that can be used with IS). That means it is difficult to use IS sampling with application that do data mining on in-memory data
- Integration Services is rather hard to use to sample data for models with nested tables. It can be done, but it takes around 11 simple steps and 14 transforms to do this for a single nested table (an example is available here: Sampling Nested Tables )
Now, there is a simpler way to do this. You may remember that, in the SQL Server Data mining architecture, a Mining Structure object acts as a data space while a mining model is a problem to be addressed in that data space. As the mining structure describes the data space, it is natural for the structure to partition the data into training and testing sets.
In SQL Server 2008, the new wizard for creating a model (or structure — yes, there is a wizard now for creating a structure with no models!) allows specifying the desired size of the “Holdout” dataset — that is, data to be stored in the mining structure for testing purposes, without being available for models training. By default, the holdout size is 30% (leaving 70% for training). You may choose to specify a fixed number of rows instead of a percentage, or both (in this case the semantic is “use 30%, but no more than 10000 rows for test data”).
The rest of this post shows how to express the holdout wizardry in DMX.
[Read more →]