Giving complexity a new credibility
InSPIRE NDA, SPSS Neural Connection and Simalto, July 1996
Tim Macer learns from people who are finding a powerful research tool in neural networks
Artificial intelligence is not something you would readily accuse your average Windows PC of. Whatever the term means, or may come to mean, AI has some way to go yet. But the program that can learn as it goes, which is one definition of AI-is available in a number of guises now, and in ways that can be of benefit to the market researcher. It is called neural networking.
Neural networking uses a complex series of mathematical models to create a decision making structure based on the way the brain works. It first runs through a lengthy learning phase, during which look for patterns in the data it is given. Once the network is fully trained, it will give you answers to questions, not by looking them up, as most software applications would, but by working it out for itself according to the rules of the network it created. As it can sit alongside classic regression techniques such as factor and cluster analysis to perform segmentation analysis, it is sometimes mistakenly thought of as being a regression technique. It is not, as it uses an entirely different method to provide its solutions.
Pulse Train produces InSPIRE NDA which is specifically aimed at carrying out what they call neural data ascription. Morton Kromann-Larsen at Gallup Denmark has been using it very selectively to merge some of their television metering with TGI data with some success. "Generally, our clients did not like using ascribed data. I'm sure it's the feeling that they are making decisions on something that is not real data. We have been ascribing data into a quarterly database. The results look quite nice on accumulated cover, but at an individual level, it is not perfect." He also explained that the method requires a lot of individual tuning, and careful monitoring to ensure that the results are meaningful and valid.
SPSS also offers a specialist neural networking module, called Neural Connection. It contains three different ways of performing neural network modelling on your data. Their Multi-layer Perceptron model parallels regression techniques and is rather like a clever weighting program. The Radial Basis and Kohonen network methods bear more resemblance to cluster analysis. You ask the network to identify groups in your data which you want to be as similar as possible, and to describe these groups.
Mark Harrison is senior analyst at Alexander Howden Group, specialists in the reinsurance market. He has been using Neural Connection on their routine work of assessing risks and also to predict the response rate to direct marketing campaigns. "I actually found it easier to use than normal regression packages. It works well against Mosaic or Acorn classifications in identifying groups. We use it as a backup tool as it helps you to pick things out of the data that you might otherwise miss." When I asked about any concerns over the methodology, he said that he felt the neural methods were, if anything, more objective and less prone to enthusiastic "fiddling". Perish the thought!
These are not tools for the lay user. Hence, the approach taken by Research for Today, who produce Simalto. Simalto is offered as a total service based around an expert neural networking system. It can perform market segmentation and trade-off analysis to a degree and level of accuracy you would not imagine possible.
Bryan Atkin, MD of Research Solutions is one of the agencies licensed to use Simalto. He explained that there are two stages to "a Simalto". Data gathering is based around a series of very complex grids which the respondent uses to perform a highly structured trade-off game. The real power of the method comes when these trade-off grids are fed into the expert system. Each simulation is tailor made for the project, hence the need for specialists who are used to working with these tools.
Sue Coyne, Director of Business & Market Research plc told me "A lot of conjoint methods can only cope with ten elements. Simalto can handle 30 or more, which means that you have room for all the secondary factors. It is the secondary, product enhancing or motivational ones which often get missed out. But it is often these secondary ones which determine the market share-if everyone is competing on the same terms."
Simalto seems to overcome the credibility issues of many neural net methods. The answers it finds tie in closely to the earlier qual. and quan. stages."Because of the visibility of the raw data, we can demonstrate that we have got the right answer," Sue Coyne explained. "And because the results are highly actionable, it helps to raise the status of the researchers within their organization."
Jane Sheehy, Customer and Product Manager at Royal Insurance spoke of their experiences with Simalto. "We use it to set customer driven service standards. It is a very objective tool and because it is customer driven it takes us away from any political debate or argument. We have benefitted from the objectivity of its findings. Once it is accepted you can make changes right back through the organization even to things like training programmes. We like it because it is simple to understand and simple to translate into action. It actually links into our real service standards." Jane Sheehy also observed: "It was useful to involve internal staff during the early stages, as this has helped to give credibility to the service standards- that they were not just made up for no apparent reason."
Barbara Clifford, Marketing Manager at UAP Provincial Insurance claimed that Simalto "certainly raised the credibility of what MR can offer - especially when MR is seen as a cinderella activity" She put this down to the high level of involvement that she said locked people in to the project. "The real benefit was when it came to the report. It showed that market research was more than just sending out questionnaires and adding up answers. Our senior management found the results were very actionable. We looked at the standards of service and the measurable differences between expectation and service delivery. We came out with a very clear set of priorities."
As markets become more sophisticated, so too must the tools to measure them. It does now seem that neural networking can offer a way to get complex without just getting confused.
Published in Research, the magazine of the Market Research Society, July 1996, Issue number 362.
© Copyright Market Research Society/Tim Macer 1997. All rights reserved. Reproduced with permission.
Copyright © 2006-2009 meaning ltd. All rights reserved.