فاطمه حسين پورگرايلي

Abstract: Today, predicting fraud in the lending process and identifying high-risk loan applicants is one of the most important and challenging issues in the banking industry. This has led to the emergence of fraud detection systems in the lending process of banks. So far, various types of fraud detection systems have been introduced, in which machine learning methods are usually used to detect fraud. In fraud detection systems based on machine learning algorithms, relatively large data sets are used to train classification models, which usually include a large number of features. Since some of these features are redundant, using all features for training and learning classification models not only does not increase its accuracy, but also increases the complexity of the model and its execution time, as well as reducing its efficiency. Therefore, as can be seen in the studies conducted in this field, the selection of important features and the identification of redundant features and their targets from high-dimensional datasets are very effective in the performance of classification models and reduce the learning time of the model and increase its classification accuracy. find Feature selection is an important and effective pre-processing technique in solving classification problems, the main purpose of which is to select an optimal subset of the entire dataset features so that it has the highest classification accuracy with the least number of features. Therefore, in this research, we propose a new approach to fraud detection in the lending process, which is based on feature selection and hybrid learning algorithms. The proposed approach consists of four main steps: In the first step, the dataset is loaded and pre-processed. In the second step, the important features are selected by a meta-heuristic optimization algorithm called Improved-BNNA-SA from the entire set of features. Then, in the third step, classification models such as support vector machines (SVM), artificial neural networks (ANNs) and decision trees are trained by the features selected in the previous step. At the end, fraud detection and identification of high-risk loan applicants is done by trained classification models. The voting method is used to aggregate the predictions obtained from the three classification models mentioned in the combined classification model. The performance of the proposed Improved-BNNA-SA algorithm for feature selection has been eva‎luated on a limited number of UCI database datasets and the results have been compared with other similar feature selection algorithms. After ensuring the better performance of the proposed algorithm, this algorithm has been applied to the bank dataset which is the main problem of this research for feature selection. The set of features selected by this algorithm has been used to train classification models, and the obtained results indicate the better performance of the proposed system to detect fraud in the lending process.

يادگيري ماشين , داده‌هاي كلان , داده‌كاوي , انتخاب ويژگي , الگوريتم‌هاي فراابتكاري , كشف تقلب , يادگيري عميق

Machine learning , big data , data mining , feature selection , meta-heuristic algorithms , fraud detection , deep learning

Fatemeh Hosseinpourgraily

Alireza Aliahmadi