bigstatsr (version 0.6.2)

big_spLogReg: Sparse logistic regression

Description

Fit lasso penalized linear regression path for a Filebacked Big Matrix. Covariates can be added to correct for confounders.

Usage

big_spLogReg(X, y01.train, ind.train = rows_along(X),
  ind.col = cols_along(X), covar.train = NULL, ncores = 1, ...)

Arguments

X

A FBM.

y01.train

Vector of responses, corresponding to ind.train. Must be only 0s and 1s.

ind.train

An optional vector of the row indices that are used, for the training part. If not specified, all rows are used. Don't use negative indices.

ind.col

An optional vector of the column indices that are used. If not specified, all columns are used. Don't use negative indices.

covar.train

Matrix of covariables to be added in each model to correct for confounders (e.g. the scores of PCA), corresponding to ind.train. Default is NULL and corresponds to only adding an intercept to each model.

ncores

Number of cores used. Default doesn't use parallelism. You may use nb_cores.

...

Arguments passed on to COPY_biglasso_main

alphas

The elastic-net mixing parameter that controls the relative contribution from the lasso (l1) and the ridge (l2) penalty. The penalty is defined as $$ \alpha||\beta||_1 + (1-\alpha)/2||\beta||_2^2.$$ alpha = 1 is the lasso penalty and alpha in between 0 (1e-4) and 1 is the elastic-net penalty. Default is 1. You can pass multiple values, and only one will be used (optimized by grid-search).

lambda.min

The smallest value for lambda, as a fraction of lambda.max. Default is .0001 if the number of observations is larger than the number of variables and .001 otherwise.

nlambda

The number of lambda values. Default is 200.

eps

Convergence threshold for inner coordinate descent. The algorithm iterates until the maximum change in the objective after any coefficient update is less than eps times the null deviance. Default value is 1e-5.

max.iter

Maximum number of iterations. Default is 1000.

dfmax

Upper bound for the number of nonzero coefficients. Default is 50e3 because, for large data sets, computational burden may be heavy for models with a large number of nonzero coefficients.

warn

Return warning messages for failures to converge and model saturation? Default is FALSE.

K

Number of sets used in the Cross-Model Selection and Averaging (CMSA) procedure. Default is 10.

ind.sets

Integer vectors of values between 1 and K specifying which set each index of the training set is in. Default randomly assigns these values.

return.all

Whether to return coefficients for all alpha and lambda values. Default is FALSE and returns only coefficients which maximize prediction on the validation sets.

nlam.min

Minimum number of lambda values to investigate. Default is 50.

n.abort

Number of lambda values for which prediction on the validation set must decrease before stopping. Default is 10.

Value

Return an object of class big_sp_best_list (a list of K elements), which has a method predict that can compute K vectors of predictions, which could be combined with e.g. rowMeans. See details.

Details

This is a modified version of one function of package biglasso. It adds the possibility to train models with covariables and use many types of FBM (not only double ones). Yet, it only corresponds to screen = "SSR" (Sequential Strong Rules).

Also, to remove the choice of the lambda parameter, we introduce the Cross-Model Selection and Averaging (CMSA) procedure:

1. This function separates the training set in K folds (e.g. 10).

2. In turn,

   • each fold is considered as an inner validation set and the others (K - 1) folds form an inner training set,

   • the model is trained on the inner training set and the corresponding predictions (scores) for the inner validation set are computed,

   • the vector of scores which maximizes log-likelihood is determined,

   • the vector of coefficients corresponding to the previous vector of scores is chosen.

3. The K resulting vectors of coefficients can then be combined into one vector (see get_beta) or you can just combine the predictions (e.g. using predict followed by rowMeans).

References

Tibshirani, R., Bien, J., Friedman, J., Hastie, T., Simon, N., Taylor, J. and Tibshirani, R. J. (2012), Strong rules for discarding predictors in lasso-type problems. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 74: 245<U+2013>266. https://doi.org/10.1111/j.1467-9868.2011.01004.x.

Zeng, Y., and Breheny, P. (2016). The biglasso Package: A Memory- and Computation-Efficient Solver for Lasso Model Fitting with Big Data in R. arXiv preprint arXiv:1701.05936. https://arxiv.org/abs/1701.05936.

See Also

glmnet biglasso

Examples

# NOT RUN {
set.seed(2)

# simulating some data
N <- 230
M <- 730
X <- FBM(N, M, init = rnorm(N * M, sd = 5))
y01 <- as.numeric((rowSums(X[, 1:5]) + 2 * rnorm(N)) > 0)
covar <- matrix(rnorm(N * 3), N)

ind.train <- sort(sample(nrow(X), 150))
ind.test <- setdiff(rows_along(X), ind.train)

test <- big_spLogReg(X, y01[ind.train], ind.train = ind.train,
                     covar.train = covar[ind.train, ])

preds <- predict(test, X, ind.row = ind.test, covar.row = covar[ind.test, ])
library(ggplot2)
qplot(preds, fill = as.logical(y01[ind.test]), geom = "density", alpha = I(0.4)) +
  labs(fill = "Case?") +
  theme_bigstatsr()
AUC(preds, y01[ind.test])
# }