Modeling heterogeneous data sets with neural networks

Abstract

This paper introduces a class of neuron models accepting heterogeneous inputs and weights. The neuron model computes a user-defined similarity function between inputs and weights. The specific similarity function used is defined by composition of a Gower-based similarity with a sigmoid function. The resulting neuron model then accepts mixtures of continuous (crisp or fuzzy) numbers, and discrete (either ordinal, integer or nominal) quantities, with explicit provision also for missing information. An artificial neural network using these neuron models is trained using a breeder genetic algorithm until convergence. A number of experiments are carried out to illustrate the validity of the approach, using several benchmarking problems. The network is compared to a standard RBF network and shown to learn from non-trivial data sets with superior generalization ability in most cases. A further advantage of the approach is the interpretability of the learned weights. © 2007 Springer-Verlag Berlin Heidelberg.