Credit assessment using constructive neural networks

Investigates the use of constructive neural networks for credit assessment. Since machine learning methods are commonly used in credit assessment tasks, the objective of this paper is to investigate the behavior of constructive neural networks, comparing their performance with that achieved by a conventional multilayer perceptron (MLP) neural network. Constructive neural networks differ from standard networks due to their ability to change their own number of elements by adding units and connections. Five constructive algorithms were used in this work: cascade correlation, tower, pyramid, upstart and M-tiling. Their main features, as well as an experiment using a credit assessment data set, are described in this work