A novel neural network technique for modelling data containing multiple functions

Abstract

Increasingly neural network techniques are being applied to a wide range of pattern recognition and classification problems. However, there is often insufficient information available to facilitate optimal operation. This problem can lead to a situation where the data exhibits signs of containing multiple underlying functions. For example, if location is not included as a feature when modelling residential property appraisal, the data will appear to map across more than one underlying function. The methodology proposed in this paper uses a form of data stratification to overcome this problem. The premise followed is that it is better to produce multiple models that are specific to — and accurate within — certain scenarios, rather than a single model that is too general and therefore inaccurate.