Although most of the studies of support vector machines (SVM) models focused on the algorithm improvement or parameters tuning, the performance of SVM models also depends on datasets, based on which the models were constructed. This paper investigate the impact of data comparability on performance of SVM models for credit scoring. After giving several examinations into data comparability and its impairing factors, we collect two practical datasets for credit scoring and then carry out several experiments to construct and test SVM models. According to the experiments'results, it has been clarified that SVM models can classify training datasets perfectly whatever data comparability may be,if we choose appropriate kernel function and related parameters. However, the performance of SVM models to classify new data depends heavily on data comparability. If data comparability is low, the accuracy for classifying test datasets is proportionally low and ?uctuates irregularly. It is obvious that guaranteeing data comparability is more important and effective than improving algorithm or turning parameters of SVM models.
展开▼
