Using Classification with K-means Clustering to Investigate Transaction Anomaly

Abstract

Applications of machine learning and related algorithms in Electronic Commerce (hereafter E-Commerce) have the potential to build robust analytical models that help examine transaction data and successfully detect and predict anomalies. Nonetheless, the robustness of such models can be undermined in the case of highly unbalanced data set. This paper presents a classification method built on K-means Clustering that addresses the issue of highly unbalanced data. In this method, we first pre-process our E-Commerce data and then apply clustering and classifying procedures to create a number of clusters where each resulting cluster includes similar transaction records. Next, four classifiers including Logistic Regression, Naive Bayes, RBFNetwork and NBtree classifiers are used to assess the resulting solution. Findings based on real-word data show that this method provides a better solution for transaction anomaly detection and prediction than traditional approaches. They also show that it straightforwardly resolves classification problems with data imbalance.