Diseño de un modelo para la detección de fraudes bancarios mediante el uso de técnicas de aprendizaje profundo

Abstract

Bank fraud is a problem worldwide, this keeps all financial institutions on alert, but thanks to the advancement of technology today, most of the entities have taken measures in this regard by implementing robust computer systems so that they do not happen. this type of deception. That is why the idea of implementing a hybrid predictive model arises to detect if a client is committing fraud based on the data that has been registered in a form for a credit card application from an Australian bank using techniques of deep learning. Within this degree work, thanks to exhaustive research, the importance of technological advances in financial companies is described, the importance of artificial intelligence, deep learning and machine learning that go hand in hand to achieve solutions for companies. such as detecting, predicting and preventing fraud by customers when filling out a form. It was proposed to develop a hybrid model that mixes two Deep Learning techniques and thanks to the bibliographical research of this work it was possible to understand what the work of a Data Analyst is in a financial entity, for this, the first thing that was done was the creation and training of a Self-organized Map (unsupervised model) that gave us as a result a list of possible fraudulent clients, from there the dependent variable was created to be able to transfer the unsupervised model to a supervised one, then we proceeded to carry out the creation and training of an Artificial Neural Network (supervised model) whose objective was to predict the percentage of whether a client is committing fraud when filling out the application for a bank credit card. To evaluate the model, the use of a comparative table was established with respect to the results of the Self-organized Map with the Original Dataset, in order to measure the percentage of precision and error. It should be noted that evaluation metrics were not used since the predictive model is not a classification model, therefore, the model does not use testing or validation datasets, nor was use of various algorithms within the Artificial Neural Network. Several performance tests of the model were carried out with variations in certain parameters of the Self-Organized Map in order to be able to choose with which parameters the best results are obtained. The best result of the tests showed a percentage of 100% accuracy of the Self-Organized Map with 18 hits and 0 errors and 97% accuracy and 15.74% loss in the training of the Artificial Neural Network. In the future, for the model to work better and more robustly, it should be experimented with larger datasets, apply more hidden layers to the artificial neural network and train it with more epochs in order to better evaluate the predictive model. It can also be experimented with by testing other types of Neural Networks depending on the data provided by the original dataset.