Usage
spa.control(eps=1e-6,maxiter=20,gcv=c("lGCV","tGCV","fGCV","aGCV"), lqmax=0.2,lqmin=0.05,ldepth=10,ltmin=0.05,lgrid=NULL, lval=NULL,dissimilar=TRUE,pce=FALSE,adjust=0,warn=FALSE,...)
Arguments
eps
the tolerance parameter for spa using a
type=class argument.
maxiter
the maximum number of iterations to run the algorithm
using type=class argument. This parameter forces the
algorithm to stop even if eps is not met.
gcv
aGCV=approximate GCV using the smoother
SLL+t(SU)*SUL, tGCV=GCV using the smoother
SLL+SLUsolve(I-SUU,SUL) (can be slow), lGCV=GCV using the
supervised smoother (fast but not that good), and fGCV=approximate GCV using the smoother S with
approximation above (this is no longer documented but it is still implemented).
lqmax
max quantile on the density of distance for data-driven estimation
lqmin
min quantile on the density of distance for data-driven estimation
ldepth
the depth of the search for divide and conquer parameter estimation
ltmin
the minimum value, in-case lqmin
is negative
lgrid
if set to an integer, then the divide and conquer approach is bypassed
lval
if set then the smoothing parameter is lval
dissimilar
if the edges represent similarity then set this to
TRUE. This flag is intended for use with the Laplacain smoother as
input (for SPS this flag is ignored and the graph is assumed to
be dissimilar). If the flag is FALSE then the supplied kernel is used to
convert the graph to similarity.
warn
if TRUE then the procedure warns the user that a ginv will
be used in the matrix inversion (i.e. the matrix is not invertible)
pce
parameter adjust is meant for adjusting hard probability
class estimates to soft (i.e. if p(z)=1 then p(z)=0.9999), for GCV
estimation, this pushes GCV away from selecting smaller values.
adjust
apply adjustment W=W+adjust.
...
mop up additional parameters passed in.