The Effect of Several Kernel Functions to Financial Transaction Anomaly Detection Performance using One-Class SVM

Anomalies in data become ubiquitous and often unavoidable. Despite it might be caused by various errors during the data collection or transportation, some anomalies potentially give important information such as indication of a new underlying process. The aim of anomaly detection task is to determine all such instances in the input dataset in a data-driven fashion. In the past decade, the proliferation of anomaly data in many domains have raised research interest resulted in a plethora of anomaly detection methods A vast number of previous study reports suggested that there is no single model will achieve the best performance for every dataset. This paper presents empiric results on the effect of several kernel functions to performance of One-Class Support Vector Machine (OC-SVM) as an anomaly detector model. The proposed model is tested using financial transactions of microfinance service dataset from an Indonesian Bank. The empiric results showed that OC-SVM model with Sigmoid and RBF kernels achieve the best statistically significant value of training, validation, and testing accuracies than the OC-SVM model with no-kernel, linear kernel and polynomial (degree 3,4,5,6) kernels.