Machine learning algorithms for financial fraud detection
Loading...
Date
2024-09-27
Authors
Μπενέκος, Κωνσταντίνος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This here study explores the effectiveness of machine learning algorithms in detecting credit card fraud. Specifically, the algorithms evaluated include random forest, logistic regression and support vector machines. The dataset utilized was
obtained from Kaggle and includes data regarding 287,807 credit card transactions which were conducted by European card holders over a two day period in September of 2013. After following standard data prepossessing procedures we continue by addressing the issue of imbalance in the data by proposing a resampling technique which entails under sampling the majority(non-fraud) class in order to create ten distinct balanced subsets of data, each of which captures a different aspect of the majority data. The models are then trained on each of the subsets separately and the final prediction made is determined by majority voting. For comparison, we also trained the models on the raw-unbalanced data in order for us to determine the effects our resampling technique had on the models overall performance. We evaluated the models both from a quantitative perspective leveraging metrics such as accuracy, recall, precision, F1 scores and most importantly ROC-AUC scores as well as in business context in the sense of how applicable/useful they would be in a real world scenario. The results revealed that while all models trained on the balanced subsets performed fairly well in terms of accuracy, recall and ROC-AUC scores they were all severely lacking in precision, something which in the fraud detection context poses a major issue. For this, it was decided that the our resampling technique negatively affected the results of our models as it made them too sensitive in predicting the fraud class which proved to be a drawback when applied to real world data. Next, when looking at the models trained on the unbalanced data we observed that our best performer in term of ROC-AUC score, which since we are talking about unbalanced classification, is our most important metric, was logistic regression with a score of 99.06%. Furthermore, logistic regression brought to the
table the added advantage of transparency, in the decision making process, as well as it not needing any prior assumption of the distribution of the data. For those reason we concluded that logistic regression arose as the most well suited modelfor fraud prediction in this particular data set.
Description
Keywords
Credit card fraud, Fraud detection, Machine learning