tune_sgspls: Compute cross-validated mean squared prediction error for...
In matt-sutton/sgspls: Sparse group-subgroup partial least squares

Description Usage Arguments Value References See Also Examples

Tuning function for finding the number of groups, and sparsities to select for an sgspls object. Offers a sequential way to find optimal sparsities, and number of groups for either block.

1 2	tune_sgspls(pls_obj, sparsities = NULL, group_seq = NULL, block = "X", folds = 10, progressBar = TRUE, setseed = 1, scale_resp = TRUE)

`pls_obj`	List of parameters or object of class `cv.sgspls` used to perform cross validation (see examples below).
`sparsities`	Matrix of sparsities, with columns corresponding to group, subgroup and individual sparsity levels to tune over. If it is NULL then a preselected set of sparsity levels is used.
`group_seq`	a vector containing the number of groups to tune over.
`block`	A string either "X" or "Y" to indicate which block to tune parameters over.
`folds`	The number of folds to use in cross validation.
`progressBar`	Logical, indicating if a progress bar is shown.
`setseed`	False, or integer for replicating tuning parameters.
`scale_resp`	Logical, the MSEP is standardised across responses (see perf function for details).

tune_sgspls returns a list with class cv.sgspls with measures:

`result_tuning`	A matrix containing the tuning parameters and MSEP values.
`best`	A vector containing the optimal tuning parameters.
`parameters`	A list of the parameters for a sgspls object.
`tuning_sparsities`	A matrix of group, subgroup and individual sparsities tuned over.
`folds`	Number of folds used in cross validation.
`min_cv`	Minimum MSEP score.
`group_seq`	Groups tuned over in cross validation.

Liquet Benoit, Lafaye de Micheaux, Boris Hejblum, Rodolphe Thiebaut. A group and Sparse Group Partial Least Square approach applied in Genomics context. Submitted.

sgspls Tuning functions calc_pve, tune_groups. Model performance and estimation predict.sgspls, perf.sgspls, coef.sgspls

 set.seed(1)
 n = 50; p = 510; 
 size.groups = 30; size.subgroups = 5
 groupX <- ceiling(1:p / size.groups)
 subgroupX <- ceiling(1:p / size.subgroups)
 
 X = matrix(rnorm(n * p), ncol = p, nrow = n)
 
 beta <- rep(0,p)
 bSG <- -2:2; b0 <- rep(0,length(bSG))
 betaG <- c(bSG, b0, bSG, b0, bSG, b0)
 beta[1:size.groups] <- betaG
 
 y = X %*% beta + rnorm(n)
 
 #--------------------------------------#
 #-- Set up a basic model to tune --#
 
 cv_pls <- list(X=X, Y=y, groupX=groupX, subgroupX=subgroupX)
 
 #---------------------------------------------#
 #-- Tune over 1 to 2 groups and multiple    --#
 #-- sparsity levels for the first component --#
 
 cv_pls_comp1 <- tune_sgspls(pls_obj = cv_pls, group_seq = 1:2, scale_resp = FALSE)
 
 #-- MSEP is on the original scale for the response --#
 cv_pls_comp1$results_tuning
 cv_pls_comp1$best
 
 ## Not run: 
 # Use the optimal fit for the first component and tune the second component
 cv_pls_comp2 <- tune_sgspls(pls_obj = cv_pls_comp1, group_seq = 1:2, scale_resp = FALSE)
 cv_pls_comp2$results_tuning
 cv_pls_comp2$best
 
 # Use the optimal fit for the second component and tune the third component
 cv_pls_comp3 <- tune_sgspls(pls_obj =  cv_pls_comp2, group_seq = 1:2, scale_resp = FALSE)
 cv_pls_comp3$best
 
 model <- do.call(sgspls, args = cv_pls_comp3$parameters)
 
 model
 
 model_coef <- coef(model, type = "coefficients")
 
 cbind(beta, model_coef$B[,,2])
 
## End(Not run)