Hello All,
I am doing LDA, trying to build a model to identify type of
bacteria. Unfortunately the training group sizes are very small....n=4 for
each group....of which there are 10.
I have noticed that LDA is highly influenced by outlying groups and
'pulls'
the LDs towards them (as you would expect).
I have looked around and found pair-wise weighted LDA. This is a rather
clever way of 'normalising' the weight given to LD scores according to
the
euclidian distance between groups, thus giving better separation power to
other groups to the minor loss of power for the outlying group.
I am able to write a programme to do this for me, but my standard results
(unweighted) are not as good as those from the R programme LDA. I know R
does LDA in a slightly more sophisticated way than me.....as I just using
bog standard (Sw-1)Sb eigenvectors due to my general ignorance.
I was wondering if anyone knew of an LDA package in R which features some
sort of 'pair-wise' distance weighting.
Thanks in advance,
Ben
--
View this message in context:
http://r.789695.n4.nabble.com/Weighted-pairwise-LDA-tp4686145.html
Sent from the R help mailing list archive at Nabble.com.