soft_impute

soft_svd

Fit SoftImpute/SoftSVD via fast alternating least squares. Based on the
paper by Trevor Hastie, Rahul Mazumder, Jason D. Lee, Reza Zadeh
by "Matrix Completion and Low-Rank SVD via Fast Alternating Least Squares" -
<a href="http://arxiv.org/pdf/1410.2596">http://arxiv.org/pdf/1410.2596</a>

Implements many algorithms for statistical learning on
sparse matrices - matrix factorizations, matrix completion,
elastic net regressions, factorization machines.
Also 'rsparse' enhances 'Matrix' package by providing methods for
multithreaded <sparse, dense> matrix products and native slicing of
the sparse matrices in Compressed Sparse Row (CSR) format.
List of the algorithms for regression problems:
1) Elastic Net regression via Follow The Proximally-Regularized Leader (FTRL)
Stochastic Gradient Descent (SGD), as per McMahan et al(, <doi:10.1145/2487575.2488200>)
2) Factorization Machines via SGD, as per Rendle (2010, <doi:10.1109/ICDM.2010.127>)
List of algorithms for matrix factorization and matrix completion:
1) Weighted Regularized Matrix Factorization (WRMF) via Alternating Least
Squares (ALS) - paper by Hu, Koren, Volinsky (2008, <doi:10.1109/ICDM.2008.22>)
2) Maximum-Margin Matrix Factorization via ALS, paper by Rennie, Srebro
(2005, <doi:10.1145/1102351.1102441>)
3) Fast Truncated Singular Value Decomposition (SVD), Soft-Thresholded SVD,
Soft-Impute matrix completion via ALS - paper by Hastie, Mazumder
et al. (2014, <doi:10.48550/arXiv.1410.2596>)
4) Linear-Flow matrix factorization, from 'Practical linear models for
large-scale one-class collaborative filtering' by Sedhain, Bui, Kawale et al
(2016, ISBN:978-1-57735-770-4)
5) GlobalVectors (GloVe) matrix factorization via SGD, paper by Pennington,
Socher, Manning (2014, <https://aclanthology.org/D14-1162/>)
Package is reasonably fast and memory efficient - it allows to work with large
datasets - millions of rows and millions of columns. This is particularly useful
for practitioners working on recommender systems.

Dmitriy Selivanov

rsparse

Statistical Learning on Sparse Matrices

David Cortes

Drew Schmidt

Wei-Chen Chen

soft_impute function

<dl><dt>x</dt>
<dd>sparse matrix. Both CSR <code>dgRMatrix</code> and CSC <code>dgCMatrix</code> are supported.
CSR matrix is preffered because in this case algorithm will benefit from multithreaded
CSR * dense matrix products (if OpenMP is supported on your platform).
On many-cores machines this reduces fitting time significantly.</dd>
<dt>rank</dt>
<dd>maximum rank of the low-rank solution.</dd>
<dt>lambda</dt>
<dd>regularization parameter for the nuclear norm</dd>
<dt>n_iter</dt>
<dd>maximum number of iterations of the algorithms</dd>
<dt>convergence_tol</dt>
<dd>convergence tolerance.
Internally functions keeps track of the relative change of the Frobenious norm
of the two consequent iterations. If the change is less than <code>convergence_tol</code>
then the process is considered as converged and function returns result.</dd>
<dt>init</dt>
<dd><a href="/link/svd?package=rsparse&version=0.5.3" data-mini-rdoc="rsparse::svd">svd</a> like object with <code>u, v, d</code> components to initialize algorithm.
Algorithm benefit from warm starts. <code>init</code> could be rank up <code>rank</code> of the maximum allowed rank.
If <code>init</code> has rank less than max rank it will be padded automatically.</dd>
<dt>final_svd</dt>
<dd><code>logical</code> whether need to make final preprocessing with SVD.
This is not necessary but cleans up rank nicely - hithly recommnded to leave it <code>TRUE</code>.</dd></dl>

Arguments

Fit SoftImpute/SoftSVD via fast alternating least squares. Based on the
paper by Trevor Hastie, Rahul Mazumder, Jason D. Lee, Reza Zadeh
by "Matrix Completion and Low-Rank SVD via Fast Alternating Least Squares" -
<a href='http://arxiv.org/pdf/1410.2596'>http://arxiv.org/pdf/1410.2596</a>

SoftImpute/SoftSVD matrix factorization — soft_impute

<dl>

<dt>x</dt>
<dd>sparse matrix. Both CSR <code>dgRMatrix</code> and CSC <code>dgCMatrix</code> are supported.
CSR matrix is preffered because in this case algorithm will benefit from multithreaded
CSR * dense matrix products (if OpenMP is supported on your platform).
On many-cores machines this reduces fitting time significantly.</dd>


<dt>rank</dt>
<dd>maximum rank of the low-rank solution.</dd>


<dt>lambda</dt>
<dd>regularization parameter for the nuclear norm</dd>


<dt>n_iter</dt>
<dd>maximum number of iterations of the algorithms</dd>


<dt>convergence_tol</dt>
<dd>convergence tolerance.
Internally functions keeps track of the relative change of the Frobenious norm
of the two consequent iterations. If the change is less than <code>convergence_tol</code>
then the process is considered as converged and function returns result.</dd>


<dt>init</dt>
<dd><a href='https://rdrr.io/r/base/svd.html'>svd</a> like object with <code>u, v, d</code> components to initialize algorithm.
Algorithm benefit from warm starts. <code>init</code> could be rank up <code>rank</code> of the maximum allowed rank.
If <code>init</code> has rank less than max rank it will be padded automatically.</dd>


<dt>final_svd</dt>
<dd><code>logical</code> whether need to make final preprocessing with SVD.
This is not necessary but cleans up rank nicely - hithly recommnded to leave it <code>TRUE</code>.</dd>

</dl>

soft_impute: SoftImpute/SoftSVD matrix factorization

Description

Usage

Value

Arguments

Examples