A Comparative Study of Distance Metrics in Machine Learning for Credit Card Fraud Detection

Abstract

Background: Credit card fraud detection is a critical problem due to the increasing volume of online transactions and the high costs associated with fraudulent activities. Previous studies in this field have investigated various machine-learning techniques to identify fraudulent transactions, with notable progress made through supervised learning methods. However, these models often face challenges due to the significant class imbalance in fraud detection datasets, where instances of fraud are much less frequent than legitimate transactions.

Purpose: As a result, there is growing interest in unsupervised techniques, such as clustering algorithms, which do not depend on labeled data and may offer improved generalization to new and unseen fraud patterns. These unsupervised approaches can autonomously identify anomalies by grouping transactions based on shared characteristics, making them a valuable alternative for detecting evolving fraudulent activities.

Methods: This work explores different distance metrics in clustering algorithms such as K-Means to identify fraudulent activity in a credit card dataset. The substantial class imbalance is highlighted by the European credit card transactions dataset, which consists of only 0.17\% of fraudulent transactions. The research utilizes multiple sampling techniques to address class imbalance.

Results: The study found that the Euclidean distance metric produced the best results out of all potential techniques when applied to the K-Means algorithm. It emphasizes how crucial it is to deal with class disparities and use unsupervised methods for fraud detection in practical settings.

Conclusions: In future research, there is scope for improvements in fraud detection systems, particularly in terms of finding enhanced algorithms and expanding data availability.