Skip to content

Suggestions for influencing the build process

There are a number of occasions when it is useful to influence how the Decision Tree grows:

  • Investigating average nodes (grey, gain approximately 1)

    These nodes may contain some responders which we could isolate by refining the rule. The node is not good enough for us to want to include our selection as is, but is not so low that we would discard the node without a further look.

  • Stopping good nodes (red, with high gain)

    These may already contain a high enough proportion of responders that we are happy to include them in the selection as is. Of course, further splitting might reveal that some people in the node are not quite as good as others, but perhaps we should concentrate effort elsewhere first.

  • Stopping bad nodes (blue, with low gain)

    These nodes contain so few responders that it is perhaps not worth our while trying to find rules to isolate them.

Controlling the build

  • Forcing a particular dimension to be used

    You can force the use of a specific dimension(s) by using the Include Dimension tick boxes on the Dimensions tab. You could then grow the whole tree using this subset of dimensions by hitting the Play button, or grow specific sections of the tree by first stopping various nodes.

  • Apply different stopping conditions to certain sections of the Decision Tree

    You could, for example, increase the Maximum Depth Stopping Condition and then build just a part of the Decision Tree. This could be useful, if there were a lot of customers in that section, such that it was worth building the tree to a deeper level.

    If a node has been stopped because the splits are not significant, you could adjust the Minimum Z-Score Stopping Condition and just grow ths node to see which nodes would have grown.

  • Influence which dimensions are used in the Decision Tree

    There may be business reasons why certain dimensions should be considered first in the Decision Tree. For example, you may want to create the first splits on age or region, if these are key business factors.

  • Experimenting with the dimensions used in the Decision Tree

    You can grow the Decision Tree from a certain point using specific dimensions or settings. Then examine the results and, if need be, prune the branch and try again with something different.