Skip to content

Choose your dimensions

The dimensions are used by the Decision Tree in creating the rules at each split. A dimension is typically a variable, although in future releases could also be a query. Variables can be dragged on to the Dimensions tab either individually or in a group or folder. They can be deleted by selecting them and using the delete key.

For an example of selecting dimensions, see Setting the dimensions.

Having selected a number of dimensions, there are various options which can be set relating to how the dimensions will be used. See How do I configure the dimensions? for more details.

Restrictions on dimensions

The following restrictions dictate which dimensions can be used:

  1. At present only selectorand flag-array variables can be used.
  2. You cannot use a dimension from a table below that which the selection is based on. The reason for this is apparent if you consider the following example, where the analysis selection is of people. Trying to use a dimension from say the holidays table could result in a person being allocated to more than one node. If the Decision Tree was to propose splitting people using booking date, with those who had a “booking date >= 2004” in one node and those with a “booking date<2004” in the other node, a person could satisfy both rules since people have more than one holiday. This is due to the one-to-many type relationships which exist between the higher and lower tables.

Considerations in choosing dimensions

In addition, the following issues should be considered:

  1. It is OK to include a variety of variables in the dimensions and allow the Decision Tree to identify which are useful. The Decision Tree will consider all the dimensions at each step and use whichever is best (according to the criteria in the Algorithm Option).
  2. Using variables that are correlated (e.g. Town and Region) is not such an issue with a Decision Tree as it is with a PWE model. If one of these variables is used for a split, the other will only be used for a later split if it helps to further identify people in the analysis selection. You can use this to your advantage, by doing early splits using high level variables such as region, and then manually forcing the use of lower level variables such as town to later investigate certain nodes in more detail.
  3. You should avoid using characteristics which are a direct result of the analysis selection. For example, if you had destination airport of Stockholm on your database and included in the model, it would inevitably be a strong characteristic of people who have been to Sweden, but would not be particularly helpful to describe these customers or predict new customers.
  4. For the model to be useful for predicting new customers, the dimensions need to be characteristics which are available on prospects.