Machine Learning by Constructing Decision Trees Including Cost Functions
Wolfgang Müller, Fritz Wysotzki · 1993
Machine Learning is now in a state to get major industrial applications. The most important application fields are diagnosis and prediction being special forms of classification. One approach to classification learning is the construction of decision trees from examples. In this case the learning algorithm constructs a decision tree from a training data set. Let the assumed training data set include n samples, each defined by a vector mj containing the attributes p1,…,p and an attached class label cj (j = 1,…,k.). This data set can be interpreted as a set of n points in an r-dimensional feature space, defined by the dimensions p1,…,p and partitioned into class regions characterised by a class label. Applications of classification learning of more static character are medical diagnosis, forecasting in banking and economy, satellite image interpretation, etc., but there is an increasing interest in application in process diagnosis and control. The classes may reflect either discretisized objective functions (e.g. intervals of high water levels as shown in the example below) or qualitative statements (e.g. the amount of the dust emitted at a power station). The below described classification algorithm CAL5 tries to discretisize automatically the feature space into areas with unique class labels. Its output is a decision tree (see figure) reflecting the resulting discretization with a high transparency. This algorithm is especially suitable for continuous and ordered discrete valued attributes Pi, respectively.