Suppose you are building a decision tree to detect fraudulent transactions. You would have a predictable column, say, [IsFraudulent], and many other input columns. This is a typical classification task - classifying new user’s behavior into two groups: Fraudulent, Not-Fraudulent. However, it may be difficult for a decision tree to learn a pattern for fraudulent transactions if there is not enough support:  for example, say only one out of a million transactions is fraudulent. To make it worse, the majority of patterns that may be found in fraudulent transactions can also be found in non-fraudulent transactions.

Here are a couple of techniques that can help you solve such problems using the Microsoft_Decision_Trees algorithm:

1.      Decreasing MINIMUM_LEAF_CASES

The default value for the MINIMUM_LEAF_CASES parameter is 10. This will keep a decision tree node with support less than 10 from being created by a split. This means that the model will never discover a distinctive pattern that appears in only less than 10 cases. You could reduce this parameter if you want to learn such patterns. However, note that reducing this number may results in a model with too many patterns that might not be interesting. In particular, if you have more than two values in the classifier (e.g., CreditCardType = {Copper, Silver, Gold, Platinum}) and you’re interested in patterns with small support only for a state (e.g., Platinum), you might want to prepare the training data with only two values (e.g., Platinum, Non-Platinum). Otherwise, the model will learn all patterns with small support for other values as well.

2.      Over-sampling

Another way to make the model learn patterns with very small support is to over-sample the patterns that you’re interested in. For instance, you could simply multiply fraudulent transactions so that it has enough support. Unlike reducing MINIMUM_LEAF_CASES, this approach won’t introduce any problem when you have more than two values in the classifier. However, this approach involves an additional data preparation step for over-sampling and the prediction probability out of the model should be adjusted accordingly (i.e., divided by the multiple).

One way to accomplish over-sampling is to use SQL Server 2005 Integration Services.  Say, you had a column in the source data named  [IsFraudulent].  You would

1. Use a conditional split on the [IsFraudulent] column to separate your cases

2. Then use a sampling transform to randomly reduce the number of non-fraudulent cases.

3. Finally, use a merge transform to bring the data back into a single dataset.