A Sampling Diagnostics Model for Neural System Training Optimization

This paper describes a hybrid-sampling model for bank fraud diagnosis, including those for multiple frauds in a banking system. The Multi-Layer Perceptron (MLP) network was used to analyze similarity, together with a statistical optimization model for sampling, to reduce the volume of used data in the diagnostics phase. The created MLP was utilized for banking transactions learning, in order to detect frauds. This neural network was tested with different configurations to improve diagnosis. The hybrid-sampling model was also employed to improve training results. The results have shown that the optimization strategy reduced the database volume and improved the learning process, presenting similar precisions to diagnose frauds detection, within this hybrid-sampling model.