SpicyMKL: DALMKL

In RMKL: Multiple Kernel Learning for Classification or Regression Problems

Description

This function conducts DALMKL for precomputed gramm matrices

Usage

SpicyMKL(K, y, loss = "hinge", C = 0.5, tolOuter = 0.01,
  tolInner = 1e-06, OuterMaxiter = 500, InnerMaxiter = 500,
  calpha = 10)

Arguments

K	The multiple kernel cube (3-d array)
y	The outome variable, must be -1/1
loss	The loss function to be used, must be either 'hinge' or 'logistic', default to be 'hinge'
C	tuning parameter for block one norm, default to be .5
tolOuter	change between to iterations is smaller than this, algorithms is considered to have converged for outer loop, default to be .01
tolInner	change between to iterations is smaller than this, algorithms is considered to have converged for inner loop, default to be .000001
OuterMaxiter	maximum number of allowed iteratons for outer loop, default to be 500
InnerMaxiter	maximum number of allowed iteratons for inner loop, default to be 500
calpha	Lagrangian parameter, default to be 10

Value

b Estimated Intercept

alpha coeffiencents of the dual of MKL

weight Estimated between kernel weight

rho Estimated within kernel weight

Examples

data(benchmark.data)
data.mkl=benchmark.data[[1]]
kernels=rep('radial',2)
sigma=c(2,1/20)
train.samples=sample(1:nrow(data.mkl),floor(0.7*nrow(data.mkl)),replace=FALSE)
degree=sapply(1:length(kernels), function(a) ifelse(kernels[a]=='p',2,0))
#Kernels.gen splts the data into a training and test set, and generates the desired kernel matrices.
#Here we generate two gaussisan kernel matrices with sigma hyperparameter 2 and 0.05
K=kernels.gen(data=data.mkl[,1:2],train.samples=train.samples,kernels=kernels,sigma=sigma,
degree=degree,scale=rep(0,length(kernels)))
C=0.05 #Cost parameter for DALMKL
K.train=K$K.train
K.test=K$K.test
# parameters set up
 ytr=data.mkl[train.samples,3]
#Converts list of kernel matrices in to an array with is appropriate for C++ code
 k.train=simplify2array(K.train)  
 k.test=simplify2array(K.test)
#Implement DALMKL with the hinge loss function
 spicy_svmb1n=SpicyMKL(K=k.train,y=ytr, loss='hinge',C=C)
 #Implement DALMKL with the hinge loss function
 spicy_logistic=SpicyMKL(K=k.train,y=ytr, loss='logistic',C=C)#' 

Example output

RMKL documentation built on May 2, 2019, 7:55 a.m.