MRG Evening – Neural Networks
Last night’s MRG Evening gave an insight into the new research developments using Neural Networks. Speaking was Alan Hendrickson of Pulse Train Technology.
Alan began by pointing out that the majority of advertising spend goes on reaching the wrong, irrelevant people; campaigns are not targeted specifically enough. Much of current research contains misleading, “dirty” data. Their solution to this problem is Neural Data Ascription. NDA fuses data to form single source databases. Simply described, the network learns about the relationships between input and output data, whereas conventional computing relies on step by step instructions.
These systems, although still experimental, are in use in some parts of Europe such as Denmark, where it is successfully working, overlaying BARB style data over their equivalent of TGI. The Burlosconi company in Italy is also experimenting with the system.
Neural networks can be applied to the problem of missing data; where no responses are available or given. In this situation, the network can be trained on existing data, focusing on the complete responses, during which the network will learn about the relationships between the questions being focused on and all the other questions in the survey. The respondents with missing data are then run through the network to create responses where data is required.
The problem of inconsistent data can similarly be tackled by training the network on the whole data set, and then applying this to each of the respondents in the survey, determining responses and marking suspect ones; these can then be replaced by the “clean” response.
In order to validate these systems, Alan showed some test results. For example, the marital status question in the TGI 1993 study was used. The survey respondents were randomly split into two groups; the first group was used to train the network and the second is then applied by ascribing questions.
Out of a sample of 2,583 respondents, 467 were actually single. The ascribed data gave 493 people; of these the network chose, 365 were actually single. A sum of 2168 of the 2583 respondents were correctly ascribed by the NDA process, or 83.9% correct.
Alan then gave demonstrations of the system in use to calculate a television schedule. Individual viewers preferences for programmes were predicted, with a similar success rate to the above example.
When it came to questions, the major interest seemed to be in the accuracy of this system compared to other methods. Steve Wilcox of BMRB, which currently fuses BARB data onto the TGI study, (one of the uses of the Neural Network suggested by Alan), wanted to know which study gave the more accurate data, and suggested that maybe Pulse Train Technology should spend some time researching this. A similar concern was voiced by Michael Stewart, who wanted to know by what percentage their system was more accurate than current static data. The unsubstantiated claims needed some definite comparing to current systems before anyone was going to be too impressed, it seemed.