Multi-view Kernel Completion
(c) Sahely Bhadra sahely@iitpkd.ac.in Jun. 1, 2016.
This version is debugged by Xiangju Qin, a postdoctoral researcher in Prof. Tero Aittokallio's group at FIMM, university of Helsinki. I convey my regards to her. April, 2019.
Details of the software are available in http://arxiv.org/abs/1602.02518
This package contains following three version of proposed multi-view kernel completion method along with supporting function and scripts
- MKCsdp
- MKCapp
- MKCemdb(ht)
Details descriare :
function [PredK,S,objective,iOutput]=MKCsdp(K,MID,para,init)
This solve MKCsdp formulation input :
K: kernels matrix size N x N x M. K(n1,n2,m) =kenrel values among n1 and n2 datapoints in m view
MID: a matrix of NxM {0,1} matrix where MID(n,m) = 1 indicates nth data is known in mth view
para: contain usedefined parameters : para.c1 Withinviewloss + para.c2 betweenviewloss + para.c3 L21norm regularisation
init : if init =1 then S is initialized by assignning all off-diagonal element with same values otherwise it S is randomly initalized
Output:
PredK =\hat(K) : predicted kernel matrices of size NxNxM
S : Inter-view similarity matrix
objective : objective values
iOutput : intermediate objective function
function [PredK,A,S,obj,iOutput] = MKCapp(K,MID,para,init)
This solve MKCapp formulation
input :
K: kernels matrix size N x N x M. K(n1,n2,m) =kenrel values among n1 and n2 datapoints in m view
MID: a matrix of NxM {0,1} matrix where MID(n,m) = 1 indicates nth data is known in mth view
para: contain usedefined parameters : para.c1 Withinviewloss + para.c2 betweenviewloss + para.c3 L21norm regularisation
init : if init =1 then A^{(m)} is initialized by non diagonal selfrepresentative matrix of K^{(m)} otherwise it A^{(m)} is randomly initalized
Output:
PredK =\hat(K) : predicted kernel matrices of size NxNxM
A : learnt reconstruction matrix
S : Inter-view similarity matrix
objective : objective values
iOutput : intermediate objective function
function [PredK,A,S,obj,iOutput] = MKCemdbht(K,MID,para,init)
This solve MKCembdht formulation
input :
K: kernels matrix size N x N x M. K(n1,n2,m) =kenrel values among n1 and n2 datapoints in m view
MID: a matrix of NxM {0,1} matrix where MID(n,m) = 1 indicates nth data is known in mth view
para: contain usedefined parameters : para.c1 Withinviewloss + para.c2 betweenviewloss + para.c3 L21norm regularisation
init : if init =1 then A^{(m)} is initialized by non diagonal selfrepresentative matrix of K^{(m)} otherwise it A^{(m)} is randomly initalized
Output:
PredK =\hat(K) : predicted kernel matrices of size NxNxM
A : learnt reconstruction matrix
S : Inter-view similarity matrix
objective : objective values
iOutput : intermediate objective function