Files
Abstract
Due to the rapid growth of e-commerce, finding a proper method to detect credit card fraud has become more important than ever. In machine learning, the two main methods of detecting credit card fraud are through supervised and unsupervised learning. Supervised learning methods are designed find patterns of known credit card transactions but are flawed due to missing novel patterns of fraud alongside missing patterns on which they haven’t yet been trained. Unsupervised learning, in contrast, can be used to find hidden patterns in unclassified data but is dependent on human intervention to understand how each feature within a dataset creates its patterns. This paper proposes a method to improve upon credit card fraud detection by applying elements of a supervised learning model to an unsupervised model. Using the decision tree classified learning algorithm, we found the importance of different components of a credit card transaction in relation to fraud. This was used to apply to a newly designed weighted kmeans algorithm that uses the importance of each feature as a weight, creating a hybrid model. This method was tested by using a dataset consisting of synthetic credit card transactions containing over one million transactions. By using this algorithm, the chance to accurately classify a transaction as fraud is improved to 98.51%, compared to the normal k-means clustering model accuracy of 87.92%--suggesting value in using a hybrid model over one singular method alone. Through this, the potential for future hybridization of unsupervised and supervised learning for credit card fraud is apparent.