Set the PWE options

The PWE algorithm creates splits by grouping together categories based on their PWE values. These options are set on the Advanced Options tab of the Build Options.

The way in which the splits are made is controlled by the Split Creation Strategy.

Split creation strategies¶

Binary splits¶

The default PWE algorithm uses a Mean Split strategy. This divides dimension values (e.g. income bands) into 2 groups according to whether their PWE is above/below the Mean PWE for the dimension. The PWE will be high whenever the Analysis % is high and low when the Analysis % is low.

Using the example of an income dimension, this will separate income bands (i.e. dimension values) which contain more Swedish holiday makers (i.e. have a high Analysis %, coloured red below) from those which have few Swedish holiday makers (i.e. low Analysis %, coloured blue below).

If you have specified an Unclassified Handling option of Keep Separate, then a third branch is created by separating off the Unclassified records from the rest of the records.

If you have specified Omit, then the Unclassified records are just removed from the branch that they would have been in.

Multi way splits¶

This strategy creates two or more splits by keeping together categories with a similar PWE score. Categories are merged together providing that their PWEs differ by less than the "Difference in PWE to Merge" value.

Another way of looking at it, is to consider all the categories sorted by descending PWE. As you move through the list of categories, a new branch is started whenever the change in PWE between consecutive categories is more than the "Difference in PWE to Merge" value.

It may not be possible to combine enough categories, such that the number of branches exceeds the maximum allowed. This occurs when categories are too different in their PWE values, such that there are too many categories which differ by more than the "Difference in PWE to Merge" value. In this case the Status in the Next Splits panel is set to "No Split" with a Reason of "Categories are too different to merge". Increasing the "Difference in PWE to Merge" value may resolve this.

Equally, categories may combine to create fewer branches than the minimum allowed. This occurs when categories are too similar in their PWE values, such that there are too many categories which differ by less than the "Difference in PWE to Merge" value. In this case the Status in the Next Splits panel is set to "No Split" with a Reason of "All categories are allocated to the same branch". Decreasing the "Difference in PWE to Merge" value may resolve this.

The following branches were created with a "Difference in PWE to Merge" value of 0.75. Branch 2 is created since the change in PWE from "20-30k" to "70-80k" is 0.98, i.e. greater than this value.

Applying the dimension constraints¶

Constraints relating to Unclassified values and Selector Branches affect the way that splits are created.

Unclassified Handling¶

If Unclassified has been set to Free-Floating is specified, then the result of the above processes are used.

If Unclassified has been set to Keep Separate, an additional split is created by separating off the unclassified category from the binary split created above. In the case of the Binary strategy, this will create a third branch

If Unclassified has been set to Omit, unclassified values are simply removed from the binary split created above

Selector Branches¶

In the case of Binary Splits, if Selector Branches has been set to Ordered or to Allow Wrap Around, then the result of the above process is taken as a starting point which is modified until the constraints are satisfied. This modification process is done using an iterative agglomerative clustering algorithm.

In the case of Multi-Way Splits, a more complex iterative approach is used. This is more time consuming, particularly with high cardinality variables.