block_tpls.RdDeveloped @ Melbourne Integrative Genomics.
block_tpls(
a,
y_ind,
ncomp = 1,
design,
scheme = "horst",
tau = 1,
m = NULL,
minv = NULL,
mode = "regression",
pairwise_scoring = "top_singular",
center = TRUE,
tol = 1e-06,
max_iter = 100,
solve_dual = TRUE,
transform_y = NULL,
face_scaling = NULL,
bpparam = NULL
)A list of tensor inputs (termed 'blocks') measured on the same
samples. The target block may instead be a vector or an n x q matrix, for a
response fixed per subject (see transform_y).
Index of "target" block in a.
The estimated number of components. ncomp must be explicitly set as an integer in tpls.
Numeric matrix of size length(x) x length(x) with values between 0 and 1; the i,j entry in the matrix indicates the strength of the relationship to be modelled between the i-th and j-th blocks. A value of 0 indicates no relationship, with 1 being the maximum. Alternatively, one can input "null" for a fully disconnected design (feature blocks are only connected to the target block at strength 1, but not to each other), or "full" for a fully connected design (target and feature blocks are all connected to each other equally at strength 1), or a single scalar value between 0 and 1 which will designate the relationship between feature blocks, with the relationships to the target block being 1. design must be explicitly set.
One of "horst", "factorial" or "centroid", "horst" by default.
Shrinkage constant to provide better estimate of covariance, introduced by Ledoit and Wolf (2004). NULL by default, but can set to "optimal" to use optimal shrinkage constants from Schafer and Strimmer (2005), or a single scalar value between 0 and 1.
A function which applies an orthogonal tensor tubal transform.
The inverse of m.
Currently supports tensor analogues of canonical, regression, and svd PLS modes. Defaults to "regression" mode.
The method on which to select the face of each tensor for each iteration of block.pls. Currently supports "top_singular" (pick based on weighted sum of top svd diagonal value within all 2-block combinatations), "sum_singular" (sum of diagonal values) and "frobenius" (top_singular but approximated with the Frobenius norm which is an upper bound). "top_singular" by default, but consider selecting "frobenius" for speed as it avoids a full svd decomposition.
If set to FALSE, the data tensor will not be centralized into Mean Deviation Form (see Mor et al. 2022). By default, the mean horizontal slice of the input tensor(s) are subtracted, so that all of the horizontal slices sum to 0, analgous to centering matrix data.
Positive scalar used as convergence criteria/tolerance during the
RGCCA iterative process. Defaults to 1e-06.
Integer specifying the maximum number of iterations to be run for each 2D RGCCA call.
For high dimensional data where n > p, use a dual formulation in the matrix RGCCA step that works with the gram matrix instead of the raw data. TRUE by default, only recommended to set to FALSE for testing purposes.
Whether to m-transform the target block with the other
blocks. NULL by default, which decides from its shape: a vector or an n x q
matrix is a response fixed per subject and is not transformed, so that every
face of the other blocks is compared with the response itself; an order-3
tensor is transformed. An untransformed target block is deflated on every
face after each component. TRUE copies a subject-fixed response over time and
transforms it; FALSE needs a target block that is the same at every time
point. See tpls.
How faces compete when each component picks one.
"frobenius" divides each face's score by the combined Frobenius norm of the
other blocks on that face, times that of the target block, so that faces
compete on a common scale; "none" compares the raw scores. It affects which
face is chosen only. NULL by default: "frobenius" for an untransformed
target block, "none" otherwise. See tpls.
A BiocParallelParam-class object
indicating the type of parallelisation. Does not have any effect if transform
functions explicitly set using m, minv.
A list containing:
The number of components used.
The input list of centered tensor blocks.
The index of the target block in a.
The design matrix used for block relationships.
The RGCCA scheme used ("horst", "factorial", "centroid").
The PLS mode used ("canonical" or "regression").
Whether the target block was m-transformed.
How faces were compared: "frobenius" or "none".
The pairwise_scoring method used.
List of loadings matrices per block (p_i x ncomp).
List of projected variates per block (n x ncomp).
Vector of iteration counts for each component.
Vector of selected tensor face indices per component.
Whether the blocks were centered.
List of the mean horizontal slices subtracted from each block, or NULL if they were not centered.
The transform pair used.
List of each X block's deflation coefficients (p_i x ncomp), NULL for the target block.
In regression mode, list of the target block's coefficients on each X block's projection (q x ncomp), NULL for the target block; NULL in canonical mode.
When design = "full", tau = 1, setting scheme to "horst",
"factorial" or "centroid" yields method SUMCOV, SSQCOV, SABSCOV respectively.