probability of classes for highly skewed dataset #59
Description
I may be slow in responding to issues till the end of June 2013.
Hello,
I am using Random Forest to obtain the probability map of classes (soft
classification instead of hard classification) which is possible by dividing
the total number of trees that vote for each class divided by the total number
of trees.
1. In my case, the samples are the pixels in a 100*360 image in which I am
looking for boundary of length 360 pixel in image. Therefore, out of 36000
samples (pixels) only 360 are belonging to class of "boundary" (class +1) and
the rest are belonging to class of "non-boundary" (class -1) which makes the
dataset highly skewed. How may I input this information to classifier? Do I
need to play with "cutoff" variable or "classwt" and how should I adjust them
to get the desire result?
2. There, are very few features for each sample(6 per pixel) since I believe
that they are all very informative features (I do not need to do feature
selection or know the importance of the features, I already did that). Do I
need to set the "mtry" variable to 6 to consider the total number of features?
I have total of 26 image of the same size in my data set and Im using a
leave-one-out procedure to produce the the probability map of classes for each
image.
Im using the precompiled package on 64 bit windows machine.
Thank you for your time and help in advance.
Saleh
Original issue reported on code.google.com by m.saleh....@gmail.com on 22 Jul 2013 at 7:11