Tuesday 18 October 2011

Data-driven segmentation: a caveat


Data-driven segmentation offers to provide hidden insights into audiences and reveal the most cost-effective ways of tailoring the offer and communications to meet their varying needs. It’s a reasonable claim. There clearly are hidden patterns in audience data and statistical techniques will, by definition, be better at identifying statistical differences than alternatives. Consequently, data driven segmentation can reasonably claim to provide segments that are ‘more different’ than alternative approaches.

However, any cluster segmentation is dependent not only on the differentness of its segments, but on the reliability of ascribing those segments to subsequent samples. These can only be based on a probability of ascribing records to the right category and depend for that reliability on the number of ‘golden questions’ used to do so.

The additional differentiation has to be offset against the ‘fuzziness’ of accurately ascribing the segments and the usability of however many ‘golden questions’ are needed for that level of accuracy.  There is potentially also another layer of ‘fuzziness’ when trying to implement actions based on the segments, since they have not been designed to align neatly with communications channels or categories of audience information. It’s easy to see how this could quickly offset the additional sharpness of distinctions from the original statistical model.

In some cases, this will be the case, in others not. But often, a combination of clusters and ‘hard’ definitions (based where possible on ‘natural’ breaks in distributions of data) will give clearer, more usable results, whilst still reflecting the differences revealed by close attention to the facts. This is the approach TGI Kantar used on the Arts Audiences Insight segmentation. Whilst it’s less ideologically rigorous than approaches that focus solely on the facts in hand, it’s often a more efficient and effective approach. Fewer 'golden questions' can mean less 'gold' expended.

No comments:

Post a Comment