sidanet: Sparse Integrative Discriminant Analysis for Multi-view...

In lasandrall/SIDA: Sparse Integrative Discriminant Analysis for Multi-view Structured Data

Sparse Integrative Discriminant Analysis for Multi-view Structured (Network) Data

Description

Performs sparse integrative disdcriminant analysis of multi-view structured (network) data to 1) obtain discriminant vectors that are associated and optimally separate subjects into different classes 2) estimate misclassification rate, and total correlation coefficient. The Laplacian of the underlying graph is used to smooth the discriminant vectors to encourage variables within a view that are connected to have a similar effect. Allows for the inclusion of other covariates which are not penalized in the algorithm. It is recommended to use cvSIDANet to choose best tuning parameter.

Usage

sidanet(Xdata=Xdata,Y=Y,myedges=myedges,myedgeweight=myedgeweight,
        Tau=Tau,withCov=FALSE,Xtestdata=NULL,Ytest=NULL,
        AssignClassMethod='Joint',plotIt=FALSE, standardize=TRUE,
        maxiteration=20,weight=0.5,thresh= 1e-03,eta=0.5,
        mynormLaplacianG=NULL)

Arguments

Xdata	A list with each entry containing training views of size n \times p_d, where d =1,...,D. Rows are samples and columns are variables. If covariates are available, they should be included as a separate view, and set as the last dataset. For binary or categorical covariates (assumes no ordering), we suggest the use of indicator variables.
Y	n \times 1 vector of class membership.
myedges	A list with each entry containing a M_d\times 2 matrix of edge information for each view. If a view has no edge information, set to 0; this will default to SIDA. If covariates are available as a view (Dth view), the edge information should be set to 0.
myedgeweight	A list with each entry containing a M_d\times 1 vector of weight information for each view. If a view has no weight information,set to 0; this will use the Laplacian of an unweighted graph. If covariates are available as a view (Dth view), the weight information should be set to 0.
Tau	d \times 1 vector of tuning parameter. It is recommended to use sidatunerange to obtain lower and upper bounds for the tuning parameters since too large a tuning parameter will result in a trivial solution vector (all zeros) and too small may result in non-sparse vectors.
withCov	TRUE or FALSE if covariates are available. If TRUE, please set all covariates as one dataset and should be the last dataset. For binary and categorical variables, use indicator matrices/vectors. Default is FALSE.
Xtestdata	A list with each entry containing testing views of size ntest \times p_d, where d =1,...,D. Rows are samples and columns are variables. The order of the list should be the same as the order for the training data, Xdata. Use if you want to predict on a testing dataset. If no Xtestdata, set to NULL.
Ytest	ntest \times 1 vector of test class membership. If no testing data provided, set to NULL.
AssignClassMethod	Classification method. Either Joint or Separate. Joint uses all discriminant vectors from D datasets to predict class membership. Separate predicts class membership separately for each dataset. Default is Joint
plotIt	TRUE or FALSE. If TRUE, produces discriminants and correlation plots. Default is FALSE
standardize	TRUE or FALSE. If TRUE, data will be normalized to have mean zero and variance one for each variable. Default is TRUE.
maxiteration	Maximum iteration for the algorithm if not converged.Default is 20.
weight	Balances separation and association. Default is 0.5.
thresh	Threshold for convergence. Default is 0.001.
eta	Balances the selection of network, and variables within network. Default is 0.5.
mynormLaplacianG	The normalized Laplacian of a graph. Set to NULL and this would be estimated using edge matrix and edge weights.

Details

The function will return several R objects, which can be assigned to a variable. To see the results, use the “$" operator.

Value

sidaneterror	Estimated classication error. If testing data provided, this will be test classification error, otherwise, training error
sidanetcorrelation	Sum of pairwise RV coefficients. Normalized to be within 0 and 1, inclusive.
hatalpha	A list of estimated sparse discriminant vectors for each view.
PredictedClass	Predicted class. If AssignClassMethod='Separate', this will be a ntest\times D matrix, with each column the predicted class for each data.

References

Sandra E. Safo, Eun Jeong Min, and Lillian Haine (2019) , Sparse Linear Discriminant Analysis for Multi-view Structured Data, submitted

See Also

cvSIDANet,sidatunerange, CorrelationPlots,DiscriminantPlots

Examples

library(SIDA)
##---- read in data
data(SIDANetDataExample)
##---- call sidanet algorithm to estimate discriminant vectors, and predict on testing data

#call sidanettunerange to get range of tuning paramater

Xdata=SIDANetDataExample[[1]]
Y=SIDANetDataExample[[2]]
Xtestdata=SIDANetDataExample[[3]]
Ytest=SIDANetDataExample[[4]]
myedges=SIDANetDataExample[[5]]
myedgeweight=SIDANetDataExample[[6]]


ngrid=10
mytunerange=sidanettunerange(Xdata,Y,ngrid,standardize=TRUE,weight=0.5,eta=0.5,
                myedges,myedgeweight)

# an example with Tau set as the lower bound
Tau=c(mytunerange$Tauvec[[1]][1], mytunerange$Tauvec[[2]][1])

#example with two views having edge weights
mysidanet=sidanet(Xdata,Y,myedges,myedgeweight,Tau,Xtestdata=Xtestdata,Ytest=Ytest)


test.error=mysidanet$sidaneterror

test.correlation=mysidanet$sidanetcorrelation

hatalpha=mysidanet$hatalpha

predictedClass=mysidanet$PredictedClass


##----plot discriminant and correlation plots

#---------Discriminant plot
mydisplot=DiscriminantPlots(Xtestdata,Ytest,mysidanet$hatalpha)

mycorrplot=CorrelationPlots(Xtestdata,Ytest,mysidanet$hatalpha)

lasandrall/SIDA documentation built on Oct. 19, 2022, 9:23 a.m.