Building neural networks

Terry Benzschawel

Although neural networks can take many forms, there are certain general activities that are typically required to construct most networks. The process of network construction is described in this chapter, which focuses in particular on the process of implementation in large financial firms. This can be exciting, but often there are difficulties that must be resolved. The process will be described in the context of building a neural network to detect credit card fraud. The project was completed for Citigroup and the network was deployed in their Fraud Early Warning (FEW) system in 1992.¹¹ I thank Margaret Trench for initiating these studies and her advice during this project. I am also grateful for her skilled management of the sensitive issues regarding public relations, internal compliance and external regulations. To the best of the author’s knowledge, this was the first neural network model deployed in production by a major banking company. Given the novelty of the technology and the theory underlying the approach, the process became complicated at times, and some of the issues are described below.

We begin with an overview of the process of neural network development, as