rsparse (version 0.3.3.2)

PureSVD: Soft-SVD decompomposition

Description

Creates matrix factorization model based on Soft-SVD. Soft SVD is very similar to truncated SVD with ability do add regularization based on nuclear norm.

Usage

PureSVD

Format

R6Class object.

Usage

For usage details see Methods, Arguments and Examples sections.

  model = PureSVD$new(rank = 10L,
                      lambda = 0,
                      init = NULL,
                      preprocess = identity,
                      ...)
  model$fit_transform(x, n_iter = 5L, ...)
  model$transform(x, ...)
  model$predict(x, k, not_recommend = x, ...)
  model$components

Methods

$new(rank = 10L, lambda = 0, init = NULL, preprocess = identity, ... )

creates matrix factorization model model with at most rank latent factors. If init is not null then initializes with provided SVD solution

$fit_transform(x, n_iter = 5L, ...)

fits model to an input user-item matrix. Returns factor matrix for users of size n_users * rank

$transform(x, ...)

Calculates user embeddings from given x user-item matrix. Result is n_users * rank matrix

$predict(x, k, not_recommend = x, ...)

predict top k item ids for users x (= column names from the matrix passed to fit_transform() method). Users features should be defined the same way as they were defined in training data - as sparse matrix of confidence values (implicit feedback) or ratings (explicit feedback). Column names (=item ids) should be in the same order as in the fit_transform().

$components

item factors matrix of size rank * n_items

Arguments

model

A PureSVD model.

x

An input sparse user-item matrix(of class dgCMatrix)

.

rank

integer - maximum number of latent factors

lambda

numeric - regularization parameter for nuclear norm

preprocess

function = identity() by default. User spectified function which will be applied to user-item interaction matrix before running matrix factorization (also applied in inference time before making predictions). For example we may want to normalize each row of user-item matrix to have 1 norm. Or apply log1p() to discount large counts.

not_recommend

sparse matrix or NULL - points which items should be excluided from recommendations for a user. By default it excludes previously seen/consumed items.

convergence_tol

numeric = -Inf defines early stopping strategy. We stop fitting when one of two following conditions will be satisfied: (a) we have used all iterations, or (b) relative change of frobenious norm of the two consequent solution is less then provided convergence_tol

...

other arguments. Not used at the moment

Examples

# NOT RUN {
data('movielens100k')
i_train = sample(nrow(movielens100k), 900)
i_test = setdiff(seq_len(nrow(movielens100k)), i_train)
train = movielens100k[i_train, ]
test = movielens100k[i_test, ]
rank = 32
lambda = 0
model = PureSVD$new(rank = rank,  lambda = lambda)
user_emb = model$fit_transform(sign(test), n_iter = 100, convergence_tol = 0.00001)
item_emb = model$components
preds = model$predict(sign(test), k = 1500, not_recommend = NULL)
mean(ap_k(preds, actual = test))
# }

Run the code above in your browser using DataCamp Workspace