| Correcting for imbalance in the data (Normalisation) | Building trees interactively | |
There are applications when the resulting decision is to be used as a model of decision making which can be automated. When this is the case then there are a number of considerations:
a)The decision making made by the tree model is dictated by the probabilities of each outcomes on the individual leafs. The probability of these outcomes being true may not be required during the automated decision making process. Instead, the dominant outcome only may want to be returned for each leaf.
b)Automated decision making does not have to incorporate all the leafs (profiles) of the tree. Certain leafs may be deemed to be not accurate enough and can therefore be excluded from the decision making. This is normally done by replacing the outcome of such groups by 'refer to human expert' or 'refer to other decision making modules'.
c)You must be able to estimate the quality of the performance which can be obtained from such a system.
d)The developer needs an estimate of the accuracy of the decision making resulting from a tree model. The accuracy of the tree as a whole can be expressed in terms of the accuracy in classifying individual outcome groups. The overall accuracy figure is the average of the individual accuracy figures, assuming that the error costs are equal for all outcome groups. Use the actual miss-classification cost for each outcome group to get a true overall cost of miss-classification.