6
$\begingroup$

Recently I have come across usage of cluster plot, which combines k-mean clustering along with PCA. The plot shows different clusters plotted using first two PCs. I have checked some of the threads (here and here) regarding the usage.

I want to know, during generating a cluster plot, does the data is clustered first and then PCA is done, or the reverse way (PCA followed by k-mean clustering)?

Because the second link says PCA is done followed clustering. But in the first link where an example is shown to generate a cluster plot, data is clustered first and then the cluster plot is generated.

Regarding interpretation, does the plot has to be interpreted as the number of clusters generated or are there any extra points to interpret?

$\endgroup$
1
  • $\begingroup$ How do you imagine cluster->PCA->plot path to yourself? $\endgroup$
    – ttnphns
    Apr 8, 2014 at 10:10

1 Answer 1

6
$\begingroup$

It is hard to see how you could do PCA on clusters; it is quite common to do PCA prior to clustering, particularly when you have a lot of variables. You can then use the PCs as variables.

You might be getting confused between a different two alternatives:

1) Do PCA on the data, then do k-means on the PCs, then plot the results

2) Do k-means on the data, do PCA on the data, then plot the clusters in terms of means on the PCs.

Both of these seem reasonable to me; the first may be better when there are many variables or when k-means on the data doesn't yield anything useful.

The former condenses the data in order to do cluster analysis. The latter condenses the data in order to visualize cluster analysis.

$\endgroup$
3
  • 1
    $\begingroup$ Thank you very much for nice explanation. I was confused between these two alternatives. Second option is what I was thinking. It is mainly for visualization. $\endgroup$ Apr 8, 2014 at 10:53
  • $\begingroup$ What did you mean by "plot[ting] the clusters in terms of means on the PCs"? I guess one could do clustering first, then do PCA, and then make a scatter-plot of the data points in PC1-PC2 coordinates, coloring each dot according to its cluster (that was identified before PCA). Seems like a feasible option to me. $\endgroup$
    – amoeba
    Dec 29, 2014 at 1:19
  • $\begingroup$ Yes that's what I meant. You could also do other pairs of PCs $\endgroup$
    – Peter Flom
    Dec 29, 2014 at 11:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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