A Case Based Reasoning System for Customer Credit Scoring: Comparative Study of Similarity Measure

Abstract

To deal with the customers’ credit assessment problem in a small company, we have developed a case-based reasoning system. The system assesses the credit score of a target customer only based on the features data which can be easily retrieved from daily transaction data stored in the database of the management information system. Since the credit score of a target customer is to be reasoned on the basis of similarity to past cases, it is very important how to evaluate properly the degree of similarity between a target customer and cases. In our previous study, the Euclidean distance was used as a similarity metric between a target customer and past cases. This paper aims at investigating the effect of similarity metrics on the performance of the proposed system. We consider six distances which are used as similarity metrics for case retrieval and case adaptation. These distances are based on weighted Manhattan distance and Euclidean distance, and the weights are calculated by using linear regression and multivariate discriminant analysis. We evaluate the distances by applying the system to solve the real credit assessment problems of the company and examining how the performance of the system depends on the choice of distances.