Archetypal Analysis: Segmenting on the Extremes

So, I’m one to generally heed the sage advice to not talk about religion or politics with friends and family (though I sometimes cannot help myself :)).  But if you’re like me, you’ve likely been forced to at least think about or possibly even divulge your political affiliation during the current election season here in the U.S.

The simple question we’re asked is:  are you Conservative, Liberal, or Moderate?

If I think about a continuum that is anchored “Liberal” on the left-hand side and “Conservative” on the right-hand side, I can answer that simple question by placing myself somewhere on that scale.

But is the question really that simple? Can you put yourself in one spot on ALL issues? Maybe you can, and that’s OK. I’m not judging. It’s possible, however, that you could place yourself more right, more center, or more left depending on the issue. You might be more Conservative on fiscal issues and more Liberal on social issues, or vice versa.  And your exact placement on the continuum might change (even if just slightly) depending on the specific fiscal or specific social issue.

My point is, it is difficult to place ourselves neatly into the same box on all issues and then say that we are the same as everyone else in that box.

Which brings me to what I want to share with you. I’ve recently worked on two projects where we explored the use of a relatively new segmentation technique called Archetypal Analysis.

Most traditional segmentation algorithms, like k-means clustering, segment folks based on centroids (or averages). The segments are identified and people are placed into one bucket where they are most similar to other folks in the bucket. Profiling of each segment is done and then we attempt to develop personas based on the profiling data.

Archetypal Analysis takes a different approach. First, rather than segment on centroids, Archetypal Analysis focuses on the extremes in the data set. In essence, the analysis identifies a number of Archetypes that can be thought of as personas. Then, Archetypal Analysis assumes each individual in the data set has some proximity to each of the Archetypes. I might be close to Archetype A characterized by certain attitudes and behaviors, a little further away from Archetype B but still somewhat close, and really far away from Archetype C. So, kind of like my political affiliation example and reality, I don’t fit neatly into one box. Rather, I’m closer to one ideal or another depending on the defining dimension (issue) in question.

For profiling purposes, we then grouped folks based on the Archetype they were in closest proximity to.  In both studies, we found much more distinctive profiles when compared to the profiles of the k-means segments.

I’m not saying that Archetypal Analysis should replace all other segmentation techniques and will always yield the most clearly distinct segments. In fact, as my friend and colleague Partha Dass blogs, MSI’s approach to segmentation is to explore multiple bases and multiple techniques to find the most differentiated segment solutions. However, in these two cases, Archetypal Analysis yielded the most distinct groupings. And the Archetype personas were remarkably clear.

Based on our experience with the technique thus far and our understanding of how it works, we believe that by looking at the current extremes of your data set we will be segmenting on emerging patterns of attitudes and behaviors. And that makes Archetypal Analysis the ideal approach for supporting product development. If you would like to discuss how Archetypal Analysis might work for you – reach out and we’ll share our thoughts.



Leave a Reply

Your email address will not be published. Required fields are marked *