Skip to content

How many nodes should I create?

There is no right answer to this question! It depends on the data that you are using and on how you are using the results from the Decision Tree. Some points to consider are listed below.

Reasons for creating a tree:

  • If you are using the Decision Tree to describe your customers (rather than to predict prospects), having to omany nodes can be too confusing.

    You will end up with lots and lots of detailed rules describing the customers in the analysis selection. Having fewer, but more general, rules can give a clearer picture.

  • If you are using the Decision Tree to predict prospects for a marketing campaign, you may be less concerned about understanding the rules and happy to use lots of complicated rules that define the very best customers.

    Care must still be taken to ensure that rules are representative (see point above) and are not too specific to the data you have used to build the Decision Tree, such that they will not apply to a different set of data.

Data volumes

  • More nodes usually mean smaller nodes and there is a danger that they will be too small to be representative.

    Lots of nodes will give you very detailed rules (e.g. people born in 1970 behave differently to those born in 1972) - these rules could be too specific and actually modelling small quirks in your data. Fewer rules, perhaps dividing your customers into a handful of age bands, would be more understandable, robust and transferable.

    The Minimum Z-Score Stopping Condition is a precautionary measure against this happening. In a future release, a holdout testing sample will also be used to verify that any rules developed do generalise to new, unseen data.

  • If you have more data, you will be able to build bigger trees.

    It will be possible to find more significant rules within a large dataset, and will be able to see differences between customers at a higher level of granularity. Again be careful not just to end up finding quirks in the data.

  • If you have more variables, you may need to build bigger trees.

    As you have more information available on customers, you should be able to find more differences between them to explain the behaviour of those in the analysis selection.

    If a lot of the variables are related, then you may find that only some of them are used.

Granularity

By creating more nodes you are able to divide up your customer base with greater granularity.

  • If you want to meet a specific marketing budget

    The top nodes from a small tree may give you the choice of targeting 70,000 people (taking say the top 2 nodes), or 130,000 people (taking say the top 3 nodes). If your budget means that 100,000 is the ideal number, you may want to break down some of the nodes further to try and get a more granular selection that lets you get closer to this figure.

  • If you are using all the leaf nodes of a particular branch node, you may as well prune the tree back to this node.

    The only point in creating additional rules to define the individual leaves is if you are going to only use some of the leaves and hence need the rules to distinguish between them.

Maximum power

  • As you add more nodes the Decision Tree will be able to separate with greater precision the people in the analysis selection from those not in the analysis selection. Each node that is split will create one node that has a lower analysis percentage than itself, and the other with a higher one. This leads to a more pronounced gains chart (more bowed) and therefore an increase in the Power statistic.
  • As the tree grows more and more, changes to the Power statistic become smaller and smaller and the incremental benefit to having more nodes becomes less and less.

  • There comes a point where the benefit of adding nodes does not warrant the added complexity of the tree. At this point the tree is big enough.

    There are some statistical approaches for determining this point, but for now subjective judgement can give a sensible cut-off.