TensorToolbox手册_原创文档.pdf

TensorToolboxfordense,sparse,anddecomposedn-wayarrays.

cp_als-ComputeaCPdecompositionofanytypeoftensor.

ALS交替最小二乘法求张量CP分解

P=CP_ALS(X,R)——计算张量X秩为R的最正确近似CP分解，P=[P.lambda,P.U]

P=CP_ALS(X,R,param,value,...)选择参数设置

tol-Toleranceondifferenceinfit{1.0e-4}

maxiters-Maximumnumberofiterations{50}

dimorder-Ordertoloopthroughdimensions{1:ndims(A)}

init-Initialguess[{random}|nvecs|cellarray]

printitn-Printfiteveryniterations;0fornoprinting{1}

[P,U0,out]=CP_ALS(...)alsoreturnsadditionaloutputthatcontainstheinputparameters.

Note:TheTheisdefinedas1-norm(X-full(P))/norm(X)andislooselytheproportionofthedata

describedbytheCPmodel,i.e.,afitof1isperfect.

%Examples:

%X=sptenrand([543],10);

%P=cp_als(X,2);

%P=cp_als(X,2,dimorder,[321]);

%P=cp_als(X,2,dimorder,[321],init,nvecs);

%U0={rand(5,2),rand(4,2),[]};%--InitialguessforfactorsofP

%[P,U0,out]=cp_als(X,2,dimorder,[321],init,U0);

%P=cp_als(X,2,out.params);%--Sameparamsaspreviousrun

交替泊松回归求张量X的非负CP分解

cp_apr-ComputenonnegativeCPwithalternatingPoissonregression.

M=CP_APR(X,R)computesanestimateofthebestrank-R

M=CP_APR(X,R,param,value,...)specifiesoptionalparametersandvalues.

Validparametersandtheirdefaultvaluesare:

tol-ToleranceontheinnerKKTviolation{1.0e-4}

maxiters-Maximumnumberofiterations{1000}

maxinneriters=Maximumnumberofinneriterations{10}

init-Initialguess[{random}|ktensor]

epsilon-parametertoavoiddividebyzero{100*eps}

kappatol-toleranceoncomplementaryslackness{100*eps}

kappa-offsettofixcomplementaryslackne