Dear All,
I am not that much into machine learning, but I know that R is wonderful
at that. My problem is the following: consider a set of N individuals
and an N by N symmetric matrix M.
This matrix stands for an interaction between individuals, i.e. the
element M[i,j]=M[j,i] is the (cumulative) amount of time that
individuals i and j spent together.
Now, the diagonal of the matrix is probably not meaningful (M[j,j] could
at most tell me for how long I have been observing every individual j,
but I would leave it out of the discussion).
In other words, for any individual j, I know all the individuals s/he
interacted with and for how long.
Finally, the individuals {j} are partitioned into 5 classes A,B,C,D,E.
The idea would be to train an algorithm on the matrix P (which is
defined exactly as M but on a subset of the set of individuals {j}) and
test its capability of predicting the classes the other individuals (for
whom I generate another interaction matrix Q) belong to.
Does anybody know how to achieve that? In a sense, I want to get a
"collision kernel" telling me how likely is an individual
characterized
by a certain set of contacts and associated contact durations to belong
to class A,B...E.
Any suggestion is welcome.
Best Regards
Lorenzo