Since consumers are diverse in terms of their experiences, attitudes, needs and expectations, companies frequently want to develop different brand messages and campaigns that are targeted to key segment(s) of their customers or prospects.

Market segmentation research is often used to identify and profile various segments in terms of:

  • Purchase decision criteria
  • Benefits desired/customer needs
  • Category behavior and attitudes
  • Brand/corporate attitudes
  • Lifestyle/psychographics
  • Demographics

Since segmentation studies typically contain a variety of questions that could be used as a basis for segmenting the market, researchers are often in a quandary as to which question battery and which clustering algorithm will best represent the population. This results in researchers either:

  • Selecting one or two question batteries and using the one clustering algorithm they prefer (or are most familiar with) when running the cluster analysis, OR
  • Generating a large number of alternate segmentation solutions (each using different questions and/or different clustering techniques) and then wading through the various solutions to subjectively select the solution that appears to “best” segment the market.

Neither of these tactics sounds as precise as we’d like – a little too much art and not quite enough science!

Fortunately, the recent emergence of “Cluster Ensembles” has enabled researchers to develop robust and reproducible segmentation solutions based on both multiple batteries of questions and alternate clustering algorithms. It is akin to a music ensemble in which different instruments each contribute to enhancing the music.

Cluster Ensembles employ a two-phase process

  • Phase 1 entails developing multiple cluster analyses that vary in terms of the clustering method employed (k-means, Hierarchical clustering using the number of clusters – ranging from 2 to 30, and the measures used. It is common to generate between 70 and 300 analyses during this phase.
  • Phase 2 clusters and groups respondents based on the cluster analyses generated in Phase 1. While different Meta-Clustering Algorithms can be used to cluster respondents, the program we use (Sawtooth Software CCEA Convergent Cluster & Ensemble Analysis), uses k–means (a distance-based algorithm) to cluster respondents and create a consensus solution.

The advantages of Cluster Ensemble include:

  • Combining groupings from alternate and dissimilar sets of variables (e.g., demographics, lifestyle batteries, desired benefits or needs)
  • Including a variety of clustering techniques when building the ensemble in Phase 1
  • Incorporating legacy clusters that are based on internal data
  • Uncovering better, more robust cluster solutions that are less sensitive to sample variations and outliers; and
  • Being able to find solutions that would not have been uncovered using a single approach.

As believers in using a more equal parts art and science, we are huge proponents of Cluster Ensemble and hope it continues to gain traction among market researchers as a primary segmentation algorithm.

Howard Firestone, VP Marketing Science