The development of artificial neural networks for the analysis of market research and electronic nose data
This thesis details research carried out into the application of unsupervised neural network and statistical clustering techniques to market research interview survey analysis. The objective of the research was to develop mathematical mechanisms to locate and quantify internal clusters within the data sets with definite commonality. As the data sets being used were binary, this commonality was expressed in terms of identical question answers. Unsupervised neural network paradigms are investigated, along with statistical clustering techniques. The theory of clustering in a binary space is also looked at. Attempts to improve the clarity of output of Self-Organising Maps (SOM) consisted of several stages of investigation culminating in the conception of the Interrogative Memory Structure (lMS). IMS proved easy to use, fast in operation and consistently produced results with the highest degree of commonality when tested against SOM, Adaptive Resonance Theory (ART!) and FASTCLUS. ARTl performed well when clusters were measured using general metrics. During the course of the research a supervised technique, the Vector Memory Array (VMA), was developed. VMA was tested against Back Propagation (BP) (using data sets provided by the Warwick electronic nose project) and consistently produced higher classification accuracies. The main advantage of VMA is its speed of operation - in testing it produced results in minutes compared to hours for the BP method, giving speed increases in the region of 100: 1.