kpca_igrad {kpcaIG}

KPCA-IG: Variables Interpretability in Kernel PCA

Description

KPCA-IG, kernel pca interpetable gradient. It is the fuction that gives the feature ranking, from the most to the least relevant variable. The ranking is obtained through the kernel’s partial derivatives computation. A score, which corresponds to the score mean among the sample points, is assigned to each input feature.

Usage

kpca_igrad(kpca_result, dim, mean_type = "arithmetic", trim_ratio = 0.1)

Arguments

kpca_result

The result of the previously obtained kernel PCA analysis.

dim

Number of kernel principal component to use for the computation of the scores. It should be less or equal to the number of component of the kPCA.

mean_type

Type of mean. Possible values are "arithmetic", "geometric", "harmonic", "median", or "trimmed". Default = "arithmetic"

trim_ratio

For mean_type == "trimmed", it is the fraction (0 to 0.5) of scores to be trimmed from each end before the mean is computed for a more robust to outliers arithmetic mean computation.

Value

A data frame containing the sorted variables and their scores sorted in decreasing order.

References

Briscik, M., Dillies, MA. & Déjean, S. Improvement of variables interpretability in kernel PCA. BMC Bioinformatics 24, 282 (2023). DOI: doi:10.1186/s12859-023-05404-y

Examples

library(WallomicsData)
library(kpcaIG)

Transcriptomics_Stems_s <- scale(Transcriptomics_Stems)

kpca_tan <-  kernelpca(as.matrix(Transcriptomics_Stems_s),
                          kernel = "tanhdot",
                          kpar = list(scale = 0.0001, offset = 0.01))

#Compute the most relevant genes based on the first two components of kpca_tan

kpca_ig_tan <- kpca_igrad(kpca_tan, dim = c(1,2))
head(kpca_ig_tan)