Modeling the Bank Client’s Behavior with LTF-C Neural Network

This paper describes an application of Local Transfer Function Classifier (LTF-C) to recognition of active and non-active bank clients, which was the problem of the 2 EUNITE Competition. LTF-C is a neural network solving classification problems. It has similar architecture as the Radial Basis Function neural network, but utilizes entirely different training algorithm. This algorithm is composed of: changing positions and sizes of reception fields of hidden neurons, insertion of new hidden neurons and removal of unnecessary ones during the training. LTF-C was chosen to solve this problem, because it had performed very well in other real-world problems, such as handwritten digit recognition, credit risk assessment or classification of breast cancer tissue. The modeling of the bank client’s behavior was performed in three stages. First, the data were preprocessed: nominal values were changed to numerical, every attribute was rescaled and transformed in order to equalize its histogram. Then, several tens of neural networks were trained. Finally, a committee of the best 14 networks was created. The paper presents also some possible directions of further research, which could lead to the increase of the usefulness and effectiveness of the system.