Métodos de detecção de fraude em cartões de crédito: um estudo comparativo

• Autor(a) principal: Piccin, Luiz Eduardo
• Data de Publicação: 2022
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/16072
• topic: Detecção de fraude; Regressão logística limitada; Regularização l1; Balanceamento e diversificação do conjunto de treinamento; CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA::ESTATISTICA::ANALISE DE DADOS; CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA::ESTATISTICA::INFERENCIA PARAMETRICA; CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA::ESTATISTICA::REGRESSAO E CORRELACAO

The arrival of the pandemic radically changed the consumption habits of goods and services, starting to occur almost exclusively in the virtual world, which in turn, in the context of fraud, has a greater number of loopholes when compared to the physical world. This increase in the number of online transactions (predominantly approved by credit card) has resulted in a greater number of frauds. From the business point of view, it is extremely important that companies are able to detect a fraudulent transaction, avoiding damage to the customer relationship and also financial losses. Usually, in the fraud detection process, there is a predictive model behind the scenes, which approaches the ideal when it presents high performance in the detection of fraudulent transactions and this extends to legitimate transactions (in technical terms, it means observing a low volume of false negatives and positives). In this work, we propose to compare the performance of two base classifiers when trained in two different architectures: the bounded version of logistic regression against its unbounded version, both with l1 regularization, using both balanced data (via k-means) and diversified (via bagging) as unbalanced data. On the k balanced and diversified training subsets to be built, the base classifiers are trained, combined by a weighted average and the final prediction is judged from this average. The comparative study is carried out in a real data scenario, in terms of AUC (Area Under the Curve) and other test statistics, such as KS (Kolmogorov-Smirnov), for example. The results obtained can also be compared with other works present in the literature