Supervised Learning with Qualitative and Mixed Attributes

Building classification tools to discriminate between good and bad credit risks is a supervised learning task which can be solved using different approaches (Graf and Nakhaeizadeh (1994)). In constructing such tools, generally, a set of training data, containing qualitative and quantitative attributes, is used to learn the discriminant rules. In real world of credit applications a lot of the available information about the customer and his behaviour of payment appears in qualitative, categorical attributes.On the other hand many approaches of supervised learning require quantitative, numerical input variables to be processed in the learning algorithms. Qualitative attributes first have to be transformed into a numerical form, before they can be used for the learning process.One very simple approach to handle that problem is to code each possible value of all qualitative categorical attributes in new, separate binary attributes. This leads to an increasing number of input variables, the learning process to build the rules gets more complicated. In particular neural networks need more time for training and often loose accuracy.In this paper we consider different scaling approaches — here the number of variables does not increase — to transform categorical into numerical attributes (Nishisato (1994)). We use them as input variables to learn the discriminant rules and develop a method of local scaling to enhance accuracy and stability of the rules. Using real world credit data, we evaluate the different approaches and compare the results.