Towards Credit-Fraud Detection via Sparsely Varying Gaussian Approximations

In today's age, financial institutions, and corporations, annually incur billions of dollars to safeguard against fraudulent activities. In this context, bank credit card frauds account for the majority of frauds. We address this problem by utilizing the Bayesian learning technique with an uncertainty score to ascertain any unwarranted positive, with a fair model accuracy by tackling unbalanced data that is common in this domain. In this direction, a sparse Gaussian classification method with the concept of pseudo inducing points has been proposed to handle the large dataset. The same has been performed using different kernels and a varying number of inducing points. The best results are obtained with the selection of RBF kernel and while considering a maximal number of inducing points. The model has been developed in a robust fashion so as to work over large financial data with an accuracy of 98.05 percent over the test dataset. Our result with low variance over the prediction indicates the robustness of the model. These results would have high applications in the context of credit fraud detection in high-class imbalanced data obtaining a high accuracy and confidence score.