An evaluation of deep learning models for chargeback Fraud detection in online games

More and more gamers are willing to pay for games. It has been estimated that the global gaming market is worth nearly US$160 billion. Chargeback services offer gamers the convenience of refund mechanisms but are often used by malicious online gamers to commit fraud, causing huge adverse impacts on the online game industry. To combat chargeback fraud, some online game providers resort to manual checking and blocking of malicious accounts, which may incur huge labor costs in the process. In this research, various deep learning models, including recurrent neural networks, long short-term memory networks, and gated recurrent units, are evaluated on their accuracy and performance in detecting malicious chargebacks in online games. In addition, traditional models, such as decision trees, k-nearest neighbors, support vector machines, and random forests, are also evaluated for comparison. The evaluation results show that the Matthews correlation coefficients of the deep learning models range between 0.84 and 0.97. In addition, the gated recurrent unit and long short-term memory network models also outperform other traditional machine learning models in the experiments in this research. Furthermore, the practical feasibility is also taken into consideration in this research by calculating the time overhead of a single transaction to determine whether there is a significant increase in time costs. Although deep learning models are less efficient than traditional machine learning models, deep learning models remain competent in minimizing losses of online game companies.