Abstract
The basis of any business is customer databases, which provide information on customer relations with the company. For example, in the field of banking services, the database stores information about the client, account number, data on financial transactions in online trading, purchased goods, their quantity, time of purchase, etc. A fraud detection is a field of data mining, which includes a set of methods for detecting fraudulent activities in the credit and financial sector, telecommunications, and other areas where illegal manipulations with customer accounts, tariff changes, etc. are possible. Typically, a fraud detection technique is based on the detection of events that do not fit into a specific pattern or behavioral pattern specific to a given business process or client that does not correspond to its patterns and trends. Analytical methods of Data Mining are widely used to build fraud detection systems: neural networks, decision trees, associative rules, sequential patterns, etc.
To solve this problem, the architecture of wavelet-neuro autoencoder and its learning algorithm are proposed. Such wavelet-neuro autoencoder has simple architecture, which is based on wavelet neuron. The tuning parameters of this architecture are not only synaptic weights and also parameters of adaptive wavelet activation functions. The tuning parameters of adaptive wavelet activation function are width, center and shape. Proposed wavelet-neuro autoencoder allows solving the fraud detection problem based on both big dataset and small dataset cases.
