Hello,
I performed PCA Analysis on the data set which contains 14 variables. After dimension reduction I have 3 new importan variables from PCA - Dim 1, Dim 2, Dim 3.
Do You know which values of reduced dimension should I take for futher analysis (for examle decision trees etc.) ? Is it eigenvalues or cos2 ?
I didn't understand your question. As I can see, you reduced 14 dimension to the 3 dimensions which contains the most variance of the data. You must used these 3 dimensions in yout further analysis, independent of the method (decision trees etc.)
Yes, but in PCA we obtain different values inside new dimensinos - eigenvalues for each dimension, cos2 for eac dimension and coordinates for each dimension and i am not sure which of them should I use for futher analysis...
Hello Sebastian,
I agree with Andre that your specific research aims aren't clear in your query. The three dimensions represent composite variables for which there will be scores (depending on which of several methods is chosen to create component scores). How you use these scores depends on the research question(s) you're trying to answer.
Eigenvalues will only convey the relative amount of variance information conveyed by each component; they will not inform as to the level of a specific case on the respective components, nor do they characterize how much or how little specific variance they carry from individual constituent variables (you could infer an average amount). As well, in the absence of rotation, dim1 will always carry more variance information than dim 2, and dim 2 more than dim 3, etc. Thus, the rank order of eigenvalues will match perfectly to the order of extraction of the components.
Coordinates (I presume you mean in three-dimensional space) will only carry information about the correspondence/relationship of component scores. If you retain the originally derived components, or do not use an oblique rotation of the PCA solution, then the geometric relationships will all be orthogonal (e.g., cosines of 90 degrees between any pair).
So, like Andre, I think you're more likely to find the component scores to be useful. If you still disagree, perhaps you could elaborate your query a bit so that RGate folks would have a better idea of what you're after.
Good luck with your work.
Thank You, I am so grateful for Your detailed answers. I am just on my first steps with PCA so maybe I don't understand properly. Thank You.
You may consider, preferably the first few columns, of your PCA/SVD [1, 2, 3 or 4 ] at the most, because most of the important features are located in the first column, and so on...
- Badges
- Science topic