Assessment of Class Balancing Techniques for Credit Card Deception Recognition

Abstract

Nowadays, rapid progress in virtual companies and the internet is hugely changing the way of trading. Since E-commerce provides international companies to trade with each other and bridges the global trading gap, both flexibility and rivalry have increased. Along with this, unlawful activities on cyber space are also growing. The readily available tools in the market make such activities easier for cyber criminals. The companies providing credit cards need to distinguish deceitful transactions so that customers can be guarded from fraudulent actions. Much work has been done in order to detect fraudulent transactions based on user profile, spend behaviour, past transactions, category wise. However, in order to detect and prevent credit card frauds in real time environment, more technical advancements and a more strategic approach considering the nature of dataset, which is used to train the model, are needed. Since the nature of credit card transaction dataset is always imbalanced, this paper intends to evaluate the efficacy of class balancing techniques for detecting credit card related deceptions. Two variations of balanced datasets are created by exploiting popular approaches like under sampling and oversampling (SMOTE). In order to evaluate the efficacy, a simple neural network with only one hidden layer has been implemented and it has been trained with balanced datasets. Our experimental results show that the model trained with dataset gives results that are more accurate.