چكيده به لاتين
Nowadays information of bank customers is getting rich, so banks can make model of their customers behavior. Traditionally, the banks evaluated their customers employing experts. Unfortunately, using this way, the manner of thinking of experts and their relations to customers, results major property damages to the banks. Non-performing loans are resulted from credential risks. In the most countries, monetary administrators have special sensitivity to the ratio of non-performing loans. If this ratio gets bigger than a limit, they try to decrease it.
In this thesis, using actual customer data, a model based on deep neural networks is developed that evaluate new customers using specifications of customers and the requested loan. For this goal, after cleaning the data, the data is normalized using a combined method. Then, a 5-layer model is designed such that the employing of various regularization strategies, the network is immune from overfitting. After the training process and optimizing various hyper parameters, the proposed model can evaluate the customers with the accuracy of 93 percent.
