How should I build a decision tree?

The ability to start and stop the growth of a Decision Tree provides great flexibility to the user. You can combine the automated statistical investigation of the data with your own business knowledge and experience.

A few things to try:

• Build it initially with a small sample (e.g. 5000 people). Look at which rules are found. These will be the really big factors driving your data.

  See Using a random Base Selection for more details.

• Build the initial branches using high level variables (e.g. Region) and then include lower level variables (e.g. Town) to investigate specific areas of the tree.

  See Suggested Uses of Dimension Settings for more details.

• Set up to use a whole variety of variables and then create a single split. Look at the ranking of dimensions within the Next Split panel to identify the most predictive variables in your data.

  You could restrict the early branches of the Decision Tree to just use these variables. This will simplify the rules created, but be careful not to miss interactions within the data by using too few variables at a time.

Refine the build¶

• Explore branches that have an average Analysis % (i.e. grey looking), to try and find some hot spots (red, high Analysis %). If none can be found you can always prune back the tree to simplify the final output.
• Build the tree to say 20 nodes with some initial settings for the stopping conditions. See which nodes get stopped and then re-grow these sections if need be with more relaxed conditions (e.g. greater depth). You may need to make the initial settings more restrictive.