| Summary: | 碩士 === 國立台北護理學院 === 資訊管理研究所 === 94 === Credit scoring models are important data analysis technique, and have been successful applied to risk management. Financial institutions usually employed credit scoring system to decide whether or not to grant credit to consumers who apply to them. Besides, credit scoring system obviously reduced time and cost. For the sake of improving model’s accuracy, many approaches have been proposed. However, unrepresentative samples are significantly reduced the accuracy of scoring model in practical application. Therefore many studies reported the usefulness of neural network in classified studies and prediction but without the explanatory decision of models. In addition, the previous studies propose mostly static credit scoring models without incrementally update models.
For this reason, to improve the shortcoming of past models that we combine hybrid clustering and case-based reasoning to propose incremental credit scoring model. The proposed model emphasizes that should be had the characteristic of accuracy, flexibility and transparent explanations. It involves three stages which are data cleaning, cluster indexed and knowledge retrieved. First, in order to enhance data quality, the hybrid clustering is used to distinguish clusters that have similar samples. Then the appropriateness of samples could be selected to construct clustered dataset. Then at cluster indexed stage, in order to increase the effectiveness of case retrieval process, we construct predictive cluster-index model by means of back propagation neural networks. At knowledge retrieved stage, CBR is made use of to construct credit retrieval model to determine whether credit applications should be granted or denied authorization. In the end, the experiments show that data cleaning really can improve model accuracy. The cluster-indexing CBR can provide better classification accuracy than no cluster-indexing CBR and traditional CBR.
|
|---|
