Kronecker product or row tensor product of two base-learners allowing for anisotropic penalties.
For the Kronecker product,
%A% works in the general case,
%A0% for the special case where
the penalty is zero in one direction.
For the row tensor product,
%Xa0% works for the special case where
the penalty is zero in one direction.
1 2 3 4 5
bl1 %A% bl2 bl1 %A0% bl2 bl1 %Xa0% bl2
base-learner 1, e.g.
base-learner 2, e.g.
%O% is called with a specification of
df in both base-learners,
bbs(x1, df = df1) %O% bbs(t, df = df2), the global
df for the
Kroneckered base-learner is computed as
df = df1 * df2.
And thus the penalty has only one smoothness parameter lambda resulting in an isotropic penalty,
P = lambda * [(P1 o I) + (I o P2)],
with overall penalty P, Kronecker product o,
marginal penalty matrices P1, P2 and identity matrices I.
(Currie et al. (2006) introduced the generalized linear array model, which has a design matrix that
is composed of the Kronecker product of two marginal design matrices, which was implemented in mboost
See Brockhaus et al. (2015) for the application of array models to functional data.)
In contrast, a Kronecker product with anisotropic penalty is obtained by
which allows for a different amount of smoothness in the two directions.
bbs(x1, df = df1) %A% bbs(t, df = df2) results in computing two
different values for lambda for the two marginal design matrices and a global value of
lambda to adjust for the global
P = lambda * [(lambda1 * P1 o I) + (I o lambda2 * P2)],
with Kronecker product o,
where lambda1 is computed individually for df1 and P1,
lambda2 is computed individually for df2 and P2,
and lambda is computed such that the global df hold df = df1 * df2.
For the computation of lambda1 and lambda2 weights specified in the model
call can only be used when the weights, are such that they are specified on the level
of rows and columns of the response matrix Y, e.g. resampling weights on the level of
rows of Y and integration weights on the columns of Y are possible.
If this the weights cannot be separated to blg1 and blg2 all
weights are set to 1 for the computation of lambda1 and lambda2 which implies that
lambda1 and lambda2 are equal over
cvrisk. The computation of the global lambda considers the
weights, such the global df are correct.
%A0% treats the important special case where lambda1 = 0 or
lambda2 = 0. In this case it suffices to compute the global lambda and computation gets
faster and arbitrary weights can be specified. Consider lambda1 = 0 then the penalty becomes
P = lambda * [(1 * P1 o I) + (I o lambda2 * P2)] = lambda * lambda2 * (I o P2),
and only one global lambda is computed which is then lambda * lambda2.
FDboost contains base-learners connected by
those effects are not expanded with
timeformula, allowing for model specifications
with different effects in time-direction.
%Xa0% computes like
%X% the row tensor product of two base-learners,
with the difference that it sets the penalty for one direction to zero.
%Xa0% behaves to
%X% analogously like
Brockhaus, S., Scheipl, F., Hothorn, T. and Greven, S. (2015): The functional linear array model. Statistical Modelling, 15(3), 279-300.
Currie, I.D., Durban, M. and Eilers P.H.C. (2006): Generalized linear array models with applications to multidimensional smoothing. Journal of the Royal Statistical Society, Series B-Statistical Methodology, 68(2), 259-280.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67
######## Example for anisotropic penalty data("viscosity", package = "FDboost") ## set time-interval that should be modeled interval <- "101" ## model time until "interval" and take log() of viscosity end <- which(viscosity$timeAll == as.numeric(interval)) viscosity$vis <- log(viscosity$visAll[,1:end]) viscosity$time <- viscosity$timeAll[1:end] # with(viscosity, funplot(time, vis, pch = 16, cex = 0.2)) ## isotropic penalty, as timeformula is kroneckered to each effect using %O% ## only for the smooth intercept %A0% is used, as 1-direction should not be penalized mod1 <- FDboost(vis ~ 1 + bolsc(T_C, df = 1) + bolsc(T_A, df = 1) + bols(T_C, df = 1) %Xc% bols(T_A, df = 1), timeformula = ~ bbs(time, df = 3), numInt = "equal", family = QuantReg(), offset = NULL, offset_control = o_control(k_min = 9), data = viscosity, control=boost_control(mstop = 100, nu = 0.4)) ## cf. the formula that is passed to mboost mod1$formulaMboost ## anisotropic effects using %A0%, as lambda1 = 0 for all base-learners ## in this case using %A% gives the same model, but three lambdas are computed explicitly mod1a <- FDboost(vis ~ 1 + bolsc(T_C, df = 1) %A0% bbs(time, df = 3) + bolsc(T_A, df = 1) %A0% bbs(time, df = 3) + bols(T_C, df = 1) %Xc% bols(T_A, df = 1) %A0% bbs(time, df = 3), timeformula = ~ bbs(time, df = 3), numInt = "equal", family = QuantReg(), offset = NULL, offset_control = o_control(k_min = 9), data = viscosity, control=boost_control(mstop = 100, nu = 0.4)) ## cf. the formula that is passed to mboost mod1a$formulaMboost ## alternative model specification by using a 0-matrix as penalty ## only works for bolsc() as in bols() one cannot specify K ## -> model without interaction term K0 <- matrix(0, ncol = 2, nrow = 2) mod1k0 <- FDboost(vis ~ 1 + bolsc(T_C, df = 1, K = K0) + bolsc(T_A, df = 1, K = K0), timeformula = ~ bbs(time, df = 3), numInt = "equal", family = QuantReg(), offset = NULL, offset_control = o_control(k_min = 9), data = viscosity, control=boost_control(mstop = 100, nu = 0.4)) ## cf. the formula that is passed to mboost mod1k0$formulaMboost ## optimize mstop for mod1, mod1a and mod1k0 ## ... ## compare estimated coefficients ## Not run: par(mfrow=c(4, 2)) plot(mod1, which = 1) plot(mod1a, which = 1) plot(mod1, which = 2) plot(mod1a, which = 2) plot(mod1, which = 3) plot(mod1a, which = 3) funplot(mod1$yind, predict(mod1, which=4)) funplot(mod1$yind, predict(mod1a, which=4)) ## End(Not run)