Generación de un conjunto de datos sintéticos mediante técnicas de aprendizaje automático para análisis de fraude

Abstract

Today, fraud-related activities are growing at a dizzying rate, causing substantial economic losses every year. For an adequate analysis of this phenomenon, it is necessary to have data that evidences this behavior, but since these are scarce and difficult to find, the gene- ration of synthetic data for its study is a viable option. Machine Learning (ML) techniques were used for the generation of text, specifically deep learning models such as Recurrent Neural Network (RNN) and Long Short Term Memory Networks (LSTM) supported by the theory of the fraud triangle proposed by Donald R. Cressey, to build a synthetic data set that allows its analysis. The RNN works with many copies of itself; each sends a message to its successor, which reduces precision when generating sentences; besides, it faces the disappearance gradient problem. The LSTM model aims to solve this problem; these networks can maintain long-term interrelationships by expanding their memory to learn from past experiences, making them perfect for generating text. The results obtained indicate that the data generation architecture proposed using the LSTM algorithm provides better sentence generation performances. The data’s readability is superior with an efficiency of 70 % compared to the RNN algorithm approach reached 40 %. Using this technique (LSTM), it was possible to synthesize a set of understandable data related to the fraud triangle to allow the fraud study to be carried out effectively.