Assessing Check Credit With Skewed Data: A Knowledge Discovery Case Study

Abstract

There have been lots of studies focusing on the improvement of performance of the whole classification in data mining, but few of them concern the relatively accuracy rate within subclasses of the whole classification. However, a real world data may exist many unproportionate subsets such that the supervised learning misleads a model into that of high performance but low interesting. In this work, we attempt to predict the customers’ credit of a checking account in a bank with skewed data and strive to fairly increase the accuracy rate of multiobject classification using neural network. Our objective is to determine a range of credit scoring which could be granted to the manger under a critical credit risk. In this paper, we analyze a real case of banking data using a series of elaborate experiments on some famous algorithms of neural network and to discuss how to decide the parameters of classes with a skewed data during training process. The result shows that the accuracy rate of classification of all classes is about 79% on average without data clean and the accuracy rates of classification of subclasses increase up to 64% with data clean. Based on the observation of experimental data, the conclusion is that this overdraft’s attribute dominates the average accuracy rate of classification while the proportion of training datasets govern the accuracy rate of classification of interesting subclasses. Finally, we infer that a nearly 15% range of noise about this real case is reasonable. Hence, a bank manager can be granted a range of 15 % for credit scoring on his duty.