Background

Despite their differences the classifiers in the previous chapter share a common ancestry – the general linear model. Consequently, they have similar relationships between the class and the predictors, albeit mediated via a link function that is determined by a defined class probability distribution. The estimated values of the predictor coefficients are found using maximum likelihood methods and tests are available to determine if they differ significantly from zero. In this chapter the methods share little in common with each other or the previous methods. Two methods are biologically inspired while a third has much in common with species identification keys. The other methods appear to work well but are less used in biology at the moment. The most important shared benefit, and potential problem, for the methods described in this chapter is that they are data-driven rather than model-driven. This can be an advantage in a knowledge-poor environment but it may come with the disadvantage that their structure is difficult to interpret. Many of these algorithms have features that reduce their susceptibility to noise and missing values, making them potentially valuable when dealing with ‘real’ data.

Decision trees

Background

Anyone who has used a species identification key should be familiar with the concept of a decision tree. Class labels are assigned to cases by following a path through a series of simple rules or questions, the answers to which determine the next direction through the pathway.