Montag, 11.5.09, 16 Uhr, Raum F 303

Fritz Wysotzki: A new, cost dependent information measure and the construction of decision trees with object dependent costs for misclassifications

It is described how costs of misclassification given with the individual training objects can be used in learning decision trees for cost optimal instead of error minimal class decisions. This is demonstrated by introducing modified, cost depending probabilities, a new, cost depending information measure and using a cost sensitive extension of the algorithm CAL5 for learning decision trees. The cost dependent information measure guarantees the selection of the (locally) next best discriminating attribute in the sequential construction of the trees. It is shown to be a cost dependent generalization of the classical information measure introduced by Shannon and may therefore be of general importance for Information Theory, Knowledge Processing and Cognitive Science since subjective evaluations can be included in information. Experiments with two artificial data sets and one application example show that this approach is more adequate than a method using class dependent costs given by experts a priori.