4
$\begingroup$

My research is a about clustering a huge data set. Right now I'm not using any technique for feature selection, because I only need 3 attributes from each row. I hope I'm making myself clear, lets say I have 1000 rows of data and in each row there are at least 15 attributes (field with different category, e.g. IP add, timestamp, numbers .. etc.) From this 15 attributes I'm only using 3 attributes. Do I need to use PCA? I'm clustering when there is a perfect match. I'm still reading about PCA, it is talking about dimensional in the data, how to determine the need to use PCA. Will the PCA technique make my cluster more accurate?

$\endgroup$
2
$\begingroup$

You can use PCA to project your data set on a lower dimensional space (reduce the dimensionality) and then use clustering there. The usual caution about preprocessing the data (normalizing them etc.) applies. Effectively, clustering then is done on "artificial" attributes, that is, linear combinations of "real" attributes. Chances are you won't find a lot of perfect matches when you do this, so you'll need some other distance function.

$\endgroup$
2
$\begingroup$

If you are only using 3 features, then doing PCA is not worth it unless you have high multicollinearity and these three variables would have some reasonable interpretation when grouped together into a composite index. The intuition is that when 2 variables used in clustering are collinear, they effectively represent the same information. But that information appears twice in the data and so it gets twice the weight of all the other variables. The second condition is probably more a matter of taste and/or the intended use of your model.

$\endgroup$

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.