An Automated Knowledge Generation Approach For Managing Credit Scoring Problems

The aim of this article is to propose an example-based intelligent approach for classifying enterprises into different categories of credit risk. The data used are of both numerical and linguistic nature. The methodology used for the rule based categorization task is the well-known inductive machine learning approach, based on entropy information. The drive for this paper was the application domain, which is very common in banking management worldwide, moreover very often turns to be a confusing and time-consuming situation. The goal is to obtain a model that correctly classifies a training sample of 130 enterprises with 76 decision variables to the predetermined classes, using a substantially less amount of attributes, trying at the same time to minimize the error rate. Data are transformed into a proper input database, and then training experiments take place in order to find the optimal settings for the training phase. After having obtained the right adjustments, the training phase initiates. The aim is to produce a comprehensible decision tree as an output, which can then be transformed to a set of simple IF/THEN rules. The specific decision tree produced from the training data, uses only 16 attributes to be formed, and is equivalent to a few comprehensible and short rules consisting of 2 to 10 premise parts. As a result, the classification task is made easier to perform and check, the amount of required data is minimized, and finally the whole process is easier to use in decision-making. Furthermore, the produced decision tree works as a knowledge generator and thus, reflects the banking organization's expertise on the application domain, represented by a handy and meaningful set of rules. Finally, the classifier could be continuously reformed, by adding to it every new credit-risk case, becoming a more and more accurate and robust classification model with time.