cw.tenfold: F-Fold cross-validation for clusterwise multiblock analyses

Description

Function to perform a F-fold cross-validation applied to clusterwise multiblock analyses. This function is usually applied to various numbers of clusters and of dimensions to select their optimal values.

Usage

cw.tenfold(Y, X, blo, option = c("none", "uniform"), G, H, FOLD = 10, INIT = 20,  method = c("mbpls", "mbpcaiv", "mbregular"), Gamma = NULL, parallel.level = c("high", "low"))

Arguments

a matrix or data frame containing the dependent variable(s)

a matrix or data frame containing the explanatory variables

blo

vector of the numbers of variables in each explanatory dataset

option

an option for the block weighting (by default, the first option is chosen): none the block weight is equal to the block inertia uniform the block weight is equal to $1/K$ for $(X_1, \dots, X_K)$ and to 1 for $X$ and $Y$

an integer giving the number of clusters

an integer giving the number of dimensions of the component-based model

FOLD

an integer giving the number of folds of the F-Fold cross-validation procedure comprised between 2 and 10 (10 by default)

INIT

an integer giving the number of initializations required for the clusterwise analysis (20 by default)

method

an option for the multiblock method to be applied (by default, the first option is chosen): mbpls multiblock Partial Least Squares is applied mbpcaiv multiblock Redundancy Analysis is applied mbregular multiblock regularized regression is applied

Gamma

a numeric value of the regularization parameter for the multiblock regularized regression comprised between 0 and 1 (NULL by default). The value (Gamma=0) leads to multiblock Redundancy Analysis and (Gamma=1) to multiblock PLS

parallel.level

Level of parallel computing, i.e. initializations are carried out simultaneously (high by default) high includes all the processing units of your computer low includes only two processing units of your computer

Value

References

Bougeard, S., Abdi, H., Saporta, G., Niang, N., Submitted, Clusterwise analysis for multiblock component methods.

Examples

Run this code

  data(simdata.red) 
  Data.X <- simdata.red[c(1:8, 21:28), 1:10]
  Data.Y <- simdata.red[c(1:8, 21:28), 11:13]
  res1   <- list()
  res2   <- list()

  ## Note that the options (INIT=2) and (parallel.level = "low") are chosen to quickly
  ## illustrate the function. 
  ## For real data, instead choose (INIT=20) to avoid local optima and (parallel.level = "high")
  ## to improve the computing speed.

  for (H in c(1:2)){
    print(paste("H=", H, sep=""))
    res1[[H]] <- cw.tenfold(Y = Data.Y, X = Data.X, blo = c(5, 5), option = "none", G = 1, H,  
      FOLD = 2, INIT = 2, method = "mbpls", Gamma = NULL, parallel.level = "low")
    res2[[H]] <- cw.tenfold(Y = Data.Y, X = Data.X, blo = c(5, 5), option = "none", G = 2, H,
       FOLD = 2, INIT = 2, method = "mbpls", Gamma = NULL, parallel.level = "low")
  }
  res1.cal <- unlist(lapply(1:2, function(x) mean(sqrt(res1[[x]]$sqrmse.cal), na.rm=TRUE))) 
  res1.val <- unlist(lapply(1:2, function(x) mean(sqrt(res1[[x]]$sqrmse.val), na.rm=TRUE)))
  res2.cal <- unlist(lapply(1:2, function(x) mean(sqrt(res2[[x]]$sqrmse.cal), na.rm=TRUE))) 
  res2.val <- unlist(lapply(1:2, function(x) mean(sqrt(res2[[x]]$sqrmse.val), na.rm=TRUE)))
  
  rmse.cal <- rbind(res1.cal, res2.cal)
  rmse.val <- rbind(res1.val, res2.val)
  rownames(rmse.cal) <- rownames(rmse.val) <- paste("G", 1:2, sep = "=")
  colnames(rmse.cal) <- colnames(rmse.val) <- paste("H", 1:2, sep = "=")

  par(mfrow=c(1,2))
  matplot(t(rmse.cal), type = "o", ylab = "RMSE of calibration", xlab = "Model dimension (H)", 
        main = "Calibration", col = c("steelblue", "darkorange"), pch = c(0, 5), lwd = c(3, 3))
        legend("center", inset = .05, legend = rownames(rmse.cal), pch = c(0, 5), lwd = c(3, 3),
        col = c("steelblue", "darkorange"), horiz = TRUE, title = "Cluster number (G)")
  matplot(t(rmse.val), type = "o", ylab = "RMSE of prediction", xlab = "Model dimension (H)", 
        main = "Prediction", col = c("steelblue", "darkorange"), pch = c(0, 5), lwd = c(3, 3))
        legend("center", inset = .05, legend = rownames(rmse.val), pch = c(0, 5), lwd = c(3, 3), 
        col = c("steelblue", "darkorange"), horiz = TRUE, title = "Cluster number (G)")