1
$\begingroup$

I am doing a segmentation project and am struggling with cluster analysis in spss right now. Could you please help me get this answered:

How do I determine the quality of the clustering in spss?

In many articles/ tutorials I've read it's advisable to run a hierarchical clustering to determine the number of clusters based on agglomeration schedule and a dendogram - and then to do k-means clustering. Let's say the 1st step results are not clear and I am hesitant between 4 and 5 cluster-solutions. I can try both with k-means method - but how do I see which one's best? Same goes for any re-running of k-means clustering procedure, since every time the output is slightly different.

Thanks a lot for any info! I would also be grateful for link to any good ready tutorials on cluster analysis in spss. What I've found so far is very random and limited... Articles on cluster analysis are not enough for me, because then I don't know how to run the tests mentioned in them in spss :-[

Cheers!

$\endgroup$
0
$\begingroup$

In my opinion, the quality of your cluster solution is inherently subjective. There's no right or wrong but rather "makes sense and is actionable for us" or "doesn't make sense, is not actionable".

Now first, what you should know is that all three clustering algorithms are affected by the order of your cases. This really raises questions about the soundness of the procedure altogether. But given that there's randomness in the result anyway, you may perhaps just as well capitalize on it as long as you're transparent about what you did and your reasoning behind it.

You can run the same clustering syntax repeatedly (say 10 times) and randomly reorder cases between runs. You can automate this with Python, an example that comes close to this is Regression over Many Dependent Variables. Now you'll get 10 different solutions, compare their quality and choose what you find best.

My personal opinion is that two pieces of information matter:

  • 1 How are the frequencies of cases over clusters? You usually don't want very small or very large clusters.
  • 2 Can you describe what the clusters mean? For this you may run means plots of the input variables with the cluster variable.

if these results make sense (they're explicable) and they're actionable, I'd say you have a good solution.

$\endgroup$
  • $\begingroup$ Thanks a lot, Ruben! I realize it's subjective. Still, I thought there might be some indicators other than "it's not making sense" that the cluster solution is not optimal. $\endgroup$ – desperate-about-statistics Jun 2 '14 at 9:03
  • $\begingroup$ Actually, I am still debating which variables to include in clustering. What would be good signs that a variable should rather be omitted, could you tell? $\endgroup$ – desperate-about-statistics Jun 2 '14 at 9:06
  • $\begingroup$ Sorry for many Q in a row, but... talking about "quality", I've just noticed a weird pattern in the way clusters are formed. Somehow the solution is always 4 clusters (the NR defined by me after reviewing a dendogram), where one is characterized by highest values (cluster centers) on all variables, one - by smallest values on all var., the other two in between. This doesn't make sense since my var.are different motivations and a person can't be highly motivated by everything or have all motivations low in my context... I guess I am facing inappropriate clustering variables selection? Thanks $\endgroup$ – desperate-about-statistics Jun 2 '14 at 9:35
  • $\begingroup$ Hey where's my upvote man? But seriously, in HCA you can ask for a "tree diagram" that suggests an 'optimum point' for the number of clusters. However, the cluster sizes should probably overrule that suggestion in my opinion. $\endgroup$ – RubenGeert Jun 2 '14 at 18:45
  • $\begingroup$ 2. Which variables should (not) be used? You can use data reduction techniques such as PCA if you have many variables that seem to measure a more limited number of traits. I believe you usually don't want highly correlated input variables so check that too. But again, the most important criterion is what's useful. $\endgroup$ – RubenGeert Jun 2 '14 at 18:49
1
$\begingroup$

Some users report getting insight from silhouette plots. You can read about them here http://en.wikipedia.org/wiki/Silhouette_(clustering).

They are available in Statistics via the STATS CLUS SIL available from the SPSS Community website (www.ibm.com/developerworks/spssdevcentral) and require the Python Essentials available through that site or with your Statistics installation materials (automatically installed with Statistics V22 or later)

$\endgroup$
  • $\begingroup$ Thank you! I thought silhouette plots are only available for the 2-step method. Good to know! $\endgroup$ – desperate-about-statistics Jun 2 '14 at 9:07

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.