Neural Network for Small Sample Sizes*

Proposer/Liason: Colleen Serafin

Johnson Controls, Inc. routinely seeks consumer feedback on product usage by having a sample of consumers rate each of a variety of products using a set of adjective rating scales. Traditionally, such data are analyzed using Factor Analysis to identify subsets of ratings scales that define common underlying factors. However, human data often contain nonlinearities that are problematic for conventional linear statistics. Therefore, we are interested in using neural networks for this purpose.

A practical drawback of neural networks is their requirement for large sample sizes. Ishihara, et al. (1995) describe and demonstrate a modified Adaptive Resonance Theory network (ART1.5-SSS) for clustering adjective data obtained from small samples of human subjects (10 in this case.) In this project we would like to have developed a neural network(s) for analyzing adjective data obtained from small samples (10-30) of human subjects. One goal of the project will be to recommend the best neural network for JCI’s purposes based upon the types of data we collect and our information requirements.

Ishihara, S., Ishihara, K., Nagamachi, M., & Matsubara, Y. (1995). An automatic builder for a Kansei Engineering expert system using self-organizing neural networks.

International Journal of Industrial Ergonomics, 15, 13-24.

*Summary prepared by Colleen Serafin.

Back To the Top