spicer: spicer

In VladoUzunangelov/SPICER: SPICER: SpicyMKL in R

Description

Port of Tomioka and Suzuki's SpicyMKL to R, expanded for multiclass and probability outputs.

Usage

spicer(K, yapp, C, opt = list())

Arguments

K	N x N x M array. the (i,j,m)-element contains the (i,j)-element of the m-th kernel gram matrix.
yapp	vector of length N with sample labels. It should be a factor for binary/multiclass classification
C	regularization parameter . Large values of C induce strong regularization. For L1 regularization C is a scalar; for elasticnet, C is a vector of length 2: C(1)|x| + C(2)x^2/2
opt	list of options which control spicer behavior: loss type of loss function:'logit' (logistic regression, log(1+exp(- f(x)y))) for classification, 'square' (square loss, 0.5*(y - f(x))^2) for regression regname type of regularization: 'l1' (default), 'elasticnet' optname optimization solver for dual variable (rho) inner minimization: 'Newton' (default), 'BFGS' stop_duality_gap TRUE/FALSE. If TRUE, Spicer employs duality gap for stopping criterion of outer loop. Default TRUE. stop_ineq_violation TRUE/FALSE. If TRUE, Spicer employs violation of inequality for stopping criterion of outer loop. Default FALSE. outer_maxiter maximum number of iterations of outer loop. (default 300) inner_maxiter maximum number of iterations of inner loop. (default 500) tol_outer tolerance of stopping criteria of outer loop. (default 0.001) tol_inner tolerance of stopping criteria of inner loop. (default tol_outer/1000) miniter minimum number of iterations of outer loop. (default 30) tol_miniter minimum relative improvement required for each outer iteration after miniter. (default 0.001) calpha increment factor of gamma: gamma^(t+1)=calpha*gamma^(t). (default 10) incl_subw TRUE/FALSE. If TRUE, Spicer includes the full NxM matrix of alpha coefficients. Default FALSE. display 1:display no progress messages, 2(default):display outer loop progress messages, 3:display inner loop progress messaages.

Value

A SPICER model with the following components:

comb_alpha

N x 1 coefficient vector.

kern_weight

1 x M kernel weight vector, scaled to sum to 1

bias

bias term

activeset

indices of kernels that are active (m : kern_weight[m] is not zero).

sorted_kern_weight

vector of non-zero kernel weights sorted by magnitude, scaled to sum to 1.

opt

list of SPICER options used in run.

history

contains history of primal objective, dual objective, number of active kernels, and duality gap.

kern_alpha

If incl_subw is TRUE, the NxM matrix of alpha coefficients.

References

• V. Uzunangelov, C. K. Wong, and J. Stuart. Highly Accurate Cancer Phenotype Prediction with AKLIMATE, a Stacked Kernel Learner Integrating Multimodal Genomic Data and Pathway Knowledge. bioRxiv, July 2020.

• Suzuki,Tomioka.SpicyMKL: a fast algorithm for Multiple Kernel Learning with thousands of kernels. Mach Learn (2011) 85:77–108

VladoUzunangelov/SPICER documentation built on July 20, 2020, 12:53 a.m.