I find that more often than expected, PCA (or maybe MDS) gets a majority of the performance of any kind of unsupervised method. If you're really interested in e

I find that more often than expected, PCA (or maybe MDS) gets a majority of the performance of any kind of unsupervised method. If you're really interested in exploring the data and methodologies, then PCA is a poor stopping point... but if you just want it to work, it's surprising how well PCA tradeoffs are good tradeoffs. All the obvious caveats apply to that whole line of thought, though.