I am trying to segment customers based on demographic, behavioral, lifestyle etc into 60-65 segments inline with Claritas Prizm segments Link1 Link2

I have 1 million records and 264 variables. Reading 68 segments in detail makes me think that it is not just a clustering technique but some rule-based segments are also there. Can anyone please suggest me what should be my approach to get 60-70 clusters as in Prizm

Is unsupervised CART Link3 will be of help here? Though I have never tried CURT before but read in stackexchange LINK4 that the process has many challenges and assumptions

  • $\begingroup$ What do you mean by "Reading 68 segments in detail"? What precisely is your question? You can apply any clustering algorithm you like after defining a metric on your attributes. Be aware that high dimensions do not behave like our low dimensional intuition would suggest, so high dimensional clustering is dubious. $\endgroup$ – Stephan Kolassa Jul 14 '18 at 18:05
  • $\begingroup$ By 68 Segments I mean the one described by Claritas (Please refer Link 1 and 2 above). My question is to understand the approach so that I will be able to create 60-70 clusters more or less describing the segments as in Claritas. I tried k-means but results are not expected. Now I am planning to create a subset of data based on some rule (such as high income etc) and then run k-means to the subset. Can I reduce dimensions using PCA then run K-means? $\endgroup$ – June Jul 14 '18 at 19:14
  • 2
    $\begingroup$ Unfortunately, we won't be able to tell you how Claritas defined their segments. Yes, you can run PCA and then apply k-means. Whether this is useful we cannot tell. $\endgroup$ – Stephan Kolassa Jul 14 '18 at 19:18
  • 1
    $\begingroup$ Actually most likely it is not meaningful. And the purpose of Claritas is to have a consistent set of groups across multiple data sets and not to cluster them differently each time. $\endgroup$ – Has QUIT--Anony-Mousse Jul 15 '18 at 8:39
  • $\begingroup$ Thanks @Anony-Mousse : Do you mean instead of clustering if I do rule based segmentation, It will give better result. What I observed from Claritas segments is that they are using 30-35 variables. I can segment my data based on rule as per definition of each Claritas's segment. Is this is the correct way of doing? When aim is to segment my data similar to the segments of Claritas. $\endgroup$ – June Jul 15 '18 at 17:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.