[R] Matrix: How create a _row-oriented_ sparse Matrix (=dgRMatrix)?

[Henrik Bengtsson]

On Wed, Apr 20, 2016 at 1:25 AM, Martin Maechler
<maechler at stat.math.ethz.ch> wrote:
>>>>>> Henrik Bengtsson <henrik.bengtsson at gmail.com>
>>>>>>     on Tue, 19 Apr 2016 14:04:11 -0700 writes:
>
>     > Using the Matrix package, how can I create a row-oriented sparse
>     > Matrix from scratch populated with some data?  By default a
>     > column-oriented one is created and I'm aware of the note that the
>     > package is optimized for column-oriented ones, but I'm only interested
>     > in using it for holding my sparse row-oriented data and doing basic
>     > subsetting by rows (even using drop=FALSE).
>
>     > Here is what I get when I set up a column-oriented sparse Matrix:
>
>     >> Cc <- Matrix(0, nrow=5, ncol=5, sparse=TRUE)
>     >> Cc[1:3,1] <- 1
>
> A general ("teaching") remark :
> The above use of Matrix() is seen in many places, and is fine
> for small matrices and the case where you only use the `[<-`
> method very few times (as above).
> Also using  Matrix()  is nice when being introduced to using the
> Matrix package.
>
> However, for efficience in non-small cases, do use
>
>    sparseMatrix()
>
> directly to construct sparse matrices.
>
>
>     >> Cc
>     > 5 x 5 sparse Matrix of class "dgCMatrix"
>
>     > [1,] 1 . . . .
>     > [2,] 1 . . . .
>     > [3,] 1 . . . .
>     > [4,] . . . . .
>     > [5,] . . . . .
>     >> str(Cc)
>     > Formal class 'dgCMatrix' [package "Matrix"] with 6 slots
>     > ..@ i       : int [1:3] 0 1 2
>     > ..@ p       : int [1:6] 0 3 3 3 3 3
>     > ..@ Dim     : int [1:2] 5 5
>     > ..@ Dimnames:List of 2
>     > .. ..$ : NULL
>     > .. ..$ : NULL
>     > ..@ x       : num [1:3] 1 1 1
>     > ..@ factors : list()
>
>     > When I try to do the analogue for a row-oriented matrix, I get a
>     > "dgTMatrix", whereas I would expect a "dgRMatrix":
>
>     >> Cr <- Matrix(0, nrow=5, ncol=5, sparse=TRUE)
>     >> Cr <- as(Cr, "dsRMatrix")
>     >> Cr[1,1:3] <- 1
>     >> Cr
>     > 5 x 5 sparse Matrix of class "dgTMatrix"
>
>     > [1,] 1 1 1 . .
>     > [2,] . . . . .
>     > [3,] . . . . .
>     > [4,] . . . . .
>     > [5,] . . . . .
>
> The reason for the above behavior has been
>
> a) efficiency.  All the subassignment ( `[<-` ) methods for
>    "RsparseMatrix" objects (of which "dsRMatrix" is a special case)
>    are implemented via  TsparseMatrix.
> b) because of the general attitude that Csparse (and Tsparse to
>    some extent) are well supported in Matrix,
>    and e.g. further operations on Rsparse matrices would *again*
>    go via T* or C* sparse ones, I had decided to keep things Tsparse.

Thanks, understanding these design decisions is helpful.
Particularly, since I consider myself a rookie when it comes to the
Matrix package.

>
> [...]
>
>     > Trying with explicit coercion does not work:
>
>     >> as(Cc, "dgRMatrix")
>     > Error in as(Cc, "dgRMatrix") :
>     > no method or default for coercing "dgCMatrix" to "dgRMatrix"
>
>     >> as(Cr, "dgRMatrix")
>     > Error in as(Cr, "dgRMatrix") :
>     > no method or default for coercing "dgTMatrix" to "dgRMatrix"
>
> The general philosophy in 'Matrix' with all the class
> hierarchies and the many specific classes has been to allow and
> foster coercing to abstract super classes,
> i.e, to  "dMatrix"  or "generalMatrix", "triangularMatrix", or
> then "denseMatrix", "sparseMatrix", "CsparseMatrix" or
> "RsparseMatrix", etc
>
> So in the above  as(*, "RsparseMatrix")   should work always.

Thanks for pointing this out (and confirming as I since discovered the
virtual RsparseMatrix class in the help).

>
>
> As a summary, in other words,  for what you want,
>
>    as(sparseMatrix(.....), "RsparseMatrix")
>
> should give you what you want reliably and efficiently.

Perfect.

>
>
>     > Am I doing some wrong here?  Or is this what means that the package is
>     > optimized for the column-oriented representation and I shouldn't
>     > really work with row-oriented ones?  I'm really only interested in
>     > access to efficient Cr[row,,drop=FALSE] subsetting (and a small memory
>     > footprint).
>
> { though you could equivalently use   Cc[,row, drop=FALSE]
>   with a CsparseMatrix Cc := t(Cr),
>   couldn't you ?
> }

Yes, I actually went ahead did that, but since the code I'm writing
supports both plain matrix:es and sparse Matrix:es, and the underlying
model operates row-by-row, I figured the code would be more consistent
if I could use row-orientation everywhere.  Not a big deal.

Thanks Martin

Henrik

>
>
> Martin Maechler  (maintainer of 'Matrix')
> ETH Zurich
>