Do you think it can be used instead of k means? I obtained a correlation with the first 2 components as they carry over 90% of the weight. Would you agree on the technique?
I think it depends on your data set and what you want to do with it. If you look at my answer to this question, you will see that it indicates groups/differences. However, it certainly doesn't prove differences - it just gives you an idea of where differences may lie.
How long would it take you run a quick k-means analysis on your data? When I have a large multivariate data set, I try many different techniques to get a handle on it.
It's possible that my background in psychological research is disguising some understanding of the broader application of PCA and K-means, but I'd say the following:
- PCA is used to reduce a set of variables to a smaller number of dimensions
- k-means is used to group cases.
For example, take a study of 1000 participants who have completed 10 different ability tests (verbal, mathematics, spatial, etc.). I could use PCA (or factor analysis) to group tests in order to identify the main underlying dimensions. I could use k-means to identify types of cases.
As a side point, I often find one approach is theoretically more interesting. If the cluster analysis is just grouping in terms of high and low on the first principal component, then I find PCA to be the more meaningful analysis, and the whole concept of clusters as an arbitrary dichotomisation (or categorisation) of a continuous variable.