Search

To increase the quality of loans provision and reduce the risk involved in this process, several credit scoring models have been developed and utilized to improve the process of assessing credit worthiness. Credit scoring is an evaluation of the risk connected with lending to clients (consumers) or an organization. The Gustafson-Kessel (GK) algorithm has become one of the most valuable tools for credit scoring. However, this algorithm demonstrates a relatively poor capability to identify a subset of features from a large dataset. Most methods that use the GK algorithm require a predefined number of clusters. This paper presents a new GK-based modified binary particle swarm optimization (MBPSO) approach to increase the classification accuracy of the GK algorithm. The proposed MBPSO consists of three parts. First, the figure of particles is utilized to determine the optimal number of clusters automatically and overcome the drawback of the GK algorithm that requires a predefined number of clusters. A subset of features is identified because the same dataset may contain influencing features or a high level of noise. The two procedures are then combined in the same optimization method to increase the classification accuracy of the GK algorithm. Second, the updating function uses velocity and position to update the next position for every particle in the swarm. Third, a kernel fuzzy clustering method (KFCM) is used as the fitness function because this function can analyze high- dimensional data. These modifications are utilized as preprocessing steps before the classification of credit data is performed. Internal measures of clustering are conducted on Australian, German, and Taiwan standard datasets that contain 690, 1,000, and 30,000 instances, respectively, with several feature properties. Results show that the GK algorithm is good at separating the data into clusters. Furthermore, the fuzzy Rand validity measures of the three credit datasets derived by using the proposed method of combining the GK algorithm with a MBPSO are greater than the values of the two other compared methods. This finding means that fuzzy partitioning (classification) is robust therefore, the risk associated with loans provision can be reduced when the proposed method is used. [ABSTRACT FROM AUTHOR]