الفهرس | Only 14 pages are availabe for public view |
Abstract In recent years, the market economic order has been negatively harmed by credit card theft, which has also damaged stakeholder, financial, and consumer interest. Billions of dollars are lost each year as a result of card fraud losses. As a result, our Thesis offers a methodology for effectively tackling fraud card detection. Recently, the dataset for card fraud transactions has been imbalanced because there are many more regular transactions than there are fraudulent ones. Before addressing the fraud issue, we must address the imbalanced data issue, which arose when the cases of one class were vastly outnumbered by those of the other class. As a result, it is difficult to classify fraud because the outcome could be biased in favor of the dominant group. To address the issue of imbalanced data, this thesis first uses various resampling approaches, such as oversampling and hybrid sampling preprocessing techniques and then addresses the issue of fraud. This thesis aims to investigate fraud detection in financial services using data analysis techniques by examining several sampling methods that produce and employ synthetic data to resample the minority class to address the uneven distribution of non-fraudulent and fraudulent classes in a dataset on credit card fraud. The goal of the thesis is to evaluate these strategies’ efficacy in the context of fraud detection, which involves a highly unbalanced dataset. In this thesis, we present four objectives for dealing with the imbalanced dataset problem. The first objective establishes that oversampling is preferable to downsampling when dealing with imbalanced datasets. In the second objective, it is demonstrated that machine learning outperforms deep learning when dealing with imbalanced datasets. The overfitting problem is addressed in the third objective, and the classification is improved. Many traditional classifiers usually fail to produce high classification performance, i.e., the majority class’s accuracy is usually significantly higher than the minority class’s accuracy. The classification criterion is often set at 0.5, which is inappropriate for an unbalanced classification. In this thesis, in the fourth objective, we also address using an effective threshold modification strategy to handle unbalanced datasets. To further improve classification performance, we extend the method by applying the oversampling preprocessing technique to the training samples. |