Re: Dataset Scales and NetDraw
|
Assuming that you are projecting to columns.
Depending on your research questions:
Is it reasonable to normalize rows by dividing each row by its sum?
Is it reasonable to standardize/normalize the columns?
To the original/transformed matrix UV you can either apply the standard
projection t(UV)*UV or compute a (dis)similarity matrix using a selected
measure on its columns - for example
? https://en.wikipedia.org/wiki/Cosine_similarity
? https://en.wikipedia.org/wiki/Canberra_distance
or some other measure from the page 29 of
? https://support.sas.com/documentation/onlinedoc/stat/131/distance.pdf
Yet another option is to compute partial
?(left, right)?contributions
l(UV) = t(UV)*b(UV) ?and ?r(UV) = t(b(UV))*UV, b(UV) is the binarized
matrix (all weights set to 1) and combine them using a selected mean.
See
https://link.springer.com/article/10.1007/s11192-020-03383-y ? ?(page 632)
https://www.sciencedirect.com/science/article/pii/S0378873321000630
The main question remains how to interpret the obtained weights in terms
of your problem?
toggle quoted message
Show quoted text
I'm giving up. I tried Method? C with Bonacich'72 adjusted to scores of 1-2 but after 2 tie removals that was no nodes left standing (big shock). My dataset consists of valued attributes ranging from 0 to 166,000. Thus, I will continue with my original
way with normalization via min-max ranging followed by binarization. It works but is not logically satisfying.
|
