4 ms·
"bread and butter of practical unsupervised learning" -- true, although I might have said "exploratory data analysis". If you can make a vector out of it someh
by mturmon 14y ago
"bread and butter of practical unsupervised learning" -- true, although I might have said "exploratory data analysis".
If you can make a vector out of it somehow, it can't hurt to try PCA. Because you don't have to figure out some fancy tailored model, or really (cough, cough) understand much about the data at all. (It sounds like I'm being sarcastic, but I'm serious -- sometimes all you want is a quick look.)
Unsupervised clustering is a similar technique.
- tel 14y agoI 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.