A fuzzy decision tree approach based on data distribution construction

Abstract

In the credit risk evaluation process, credit classification is one of the most popular method to assess the according risk. In real world application, credit classification can distinguish the bad credit transactions from the good ones, which avoids potential risks. To identify the potential risks (bad credit), we employ a fuzzy decision tree method to accomplish the classification process for the credit risk evaluation, because it owns the advantages of ease of interpretation, reduction of information loss and competitive theoretical basis. To further extend these advantages, this paper proposes a novel data-distribution based fuzzy decision tree (DDBFDT) approach, which not only represents a new fuzzy partition points finding algorithm but also formulates the process of building membership function. The merits of our approach is fourfold: 1) a new error-related metric to select the best attribute for creating our fuzzy decision tree; 2) a data-distribution aware algorithm in the process of partition point searching 3) development of less computing complex non-linear membership function and more interpretable fuzzy sets building strategy with original data distribution involved; 4) better performance than similar models and high readability due to our fuzzy result of classification, as well as robustness and resisting disturbance. Our proposed DDBFDT approach is testified by both rigorous theoretics and a couple of experiments using public opened data-sets as well as synthesized data-sets.