iNMF_lambda_tuning: Integrative NMF Tune Lambda

In hdsu-bioquant/ButchR: Non-Negative Matrix Factorization Suite

Integrative NMF Tune Lambda

Description

Based on Yang and Michailidis, 2016, to tune the value of the parameter lambda for the integrative NMF (iNMF), the objectives values of join NMF (jNMF) are compared to single view NMFs (sNMF), the principle is that join NMF represents complete homogeneity and single view NMF represents complete heterogeneity. To avoid overfitting the best lambda can be selected by plotting the difference of the unsquared residual quantities of jNMF and iNMF (Ri - Rj) over multiple values of lambda, and compare it to the difference of the unsquared residual quantities of sNMF and jNMF c*(Rj - Rs). The optimal lambda usually is the first lambda in which (Ri - Rj) < c*(Rj - Rs). Where c is a constant >= 2.

Usage

iNMF_lambda_tuning(
  matrix_list,
  lambdas = seq(0, 1, 0.1),
  Output_type = "residuals",
  thr_cons = 3,
  rank = 2,
  n_initializations = 10,
  iterations = 10^4,
  convergence_threshold = 40,
  Sp = 0,
  show_plot = TRUE,
  extract_features = FALSE
)

Arguments

matrix_list	List of non-negative matrices.
lambdas	a sequence of lambdas to test.
Output_type	Type of desired output, could be: residuals - Residual quantities data.frame. iNMF - Returns the integrative NMF object for the optimal estimated lambda. all_iNMF - Returns a list with an integrative NMF object for every lambda. plot - Residual quantities plot, (Ri - Rj) and 2(Rj - Rs) vs lamda range. all - list containing all computed NMF objects, Residual quantities data.frame and Residual quantities plot.
thr_cons	numeric value, Threshold constant c, in c*(Rj - Rs).
rank	numeric vector with rank to factorize.
n_initializations	Number of initializations to evaluate.
iterations	Maximum number of iterations to run for every initialization.
convergence_threshold	The factorization stops, if the convergence test is constant for this number of iterations.
Sp	Sparcity constrain, values > 0 force sparcity in the H matrix.
show_plot	TRUE print plot; FALSE plot is only returned if it is selected in Output_type.
extract_features	TRUE Extract signature specific features.

Value

There are five different types of possible outputs, depending on the selected option on 'Output_type'. The default output is a data.frame with the residual resulting from tuning lambdas across multiple lambdas.

Examples

## Not run: 
# Compare the residual across multiple lambdas:
iNMF_lambda_tuning(matrix_list           = norm_mat_list,
                   lambdas               = seq(0, 1, 0.1),
                   Output_type           = "residuals",
                   rank                  = 9,
                   n_initializations     = 5,
                   iterations            = 10^4,
                   convergence_threshold = 40,
                   Sp                    = 0,
                   extract_features      = FALSE)
# Retrieve all the objects and extract the iNMF for the best lambda:
inmf_tune <- iNMF_lambda_tuning(matrix_list           = norm_mat_list,
                                lambdas               = seq(0, 1, 0.1),
                                thr_cons              = 4,
                                Output_type           = "all",
                                rank                  = 9,
                                n_initializations     = 5,
                                iterations            = 10^4,
                                convergence_threshold = 40)
min(inmf_tune$residuals$lambda[inmf_tune$residuals$best_lambda])
inmf_tune$iNMF$lambda_0.2

## End(Not run)

hdsu-bioquant/ButchR documentation built on Jan. 28, 2023, 6:06 p.m.