phenoRegressor.SVR: Support Vector Regression using package e1071

In GROAN: Genomic Regression Workbench

Support Vector Regression using package e1071

Description

This is a wrapper around several functions from e1071 package (as such, it won't work if e1071 package is not installed). This function implements Support Vector Regressions, meaning that the data points are projected in a transformed higher dimensional space where linear regression is possible.

phenoRegressor.SVR can operate in three modes: run, train and tune.
In run mode you need to pass the function an already tuned/trained SVR model, typically obtained either directly from e1071 functions (e.g. from svm, best.svm and so forth) or from a previous run of phenoRegressor.SVR in a different mode. The passed model is applied to the passed dataset and predictions are returned.
In train mode a SVR model will be trained on the passed dataset using the passed hyper parameters. The trained model will then be used for predictions.
In tune mode you need to pass one or more sets of hyperparameters. The best combination of hyperparameters will be selected through crossvalidation. The best performing SVR model will be used for final predictions. This mode can be very slow.

There is no distinction between regular covariates (genotypes) and extra covariates (fixed effects) in Support Vector Regression. If extra covariates are passed, they are put together with genotypes, side by side. Same thing happens with covariances matrix. This can bring to the scientifically questionable but technically correct situation of regressing on a big matrix made of SNP genotypes, covariances and other covariates, all collated side by side. The function makes no distinction, and it's up to the user understand what is correct in each specific experiment.

Usage

phenoRegressor.SVR(
  phenotypes,
  genotypes,
  covariances,
  extraCovariates,
  mode = c("tune", "train", "run"),
  tuned.model = NULL,
  scale.pheno = TRUE,
  scale.geno = FALSE,
  ...
)

Arguments

phenotypes	phenotypes, a numeric array (n x 1), missing values are predicted
genotypes	SNP genotypes, one row per phenotype (n), one column per marker (m), values in 0/1/2 for diploids or 0/1/2/...ploidy for polyploids. Can be NULL if covariances is present.
covariances	square matrix (n x n) of covariances. Can be NULL if genotypes is present.
extraCovariates	extra covariates set, one row per phenotype (n), one column per covariate (w). If NULL no extra covariates are considered.
mode	this parameter decides what will happen with the passed dataset mode = "tune" : hyperparameters will be tuned on a grid (you may want to specify its values using extra params) with a call to e1071::tune.svm. Use this option if you have no idea about the optimal choice of hyperparameters. This mode can be very slow. mode = "train" : an SVR will be trained on the train dataset using the passed hyperparameters (if you know them). This more invokes e1071::train mode = "run" : you already have a tuned and trained SVR (put it into tuned.model) and want to use it. The fastest mode.
tuned.model	a tuned and trained SVR to be used for prediction. This object is only used if mode is equal to "run".
scale.pheno	if TRUE (default) the phenotypes will be scaled and centered (before tuning or before applying the passed tuned model).
scale.geno	if TRUE the genotypes will be scaled and centered (before tuning or before applying the passed tuned model. It is usually not a good idea, since it leads to worse results. Defaults to FALSE.
...	all extra parameters are passed to e1071::svm or e1071::tune.svm

Value

The function returns a list with the following fields:

• predictions : an array of (n) predicted phenotypes

• hyperparams : named vector with the following keys: gamma, cost, coef0, nu, epsilon. Some of the values may not make sense given the selected model, and will contain default values from e1071 library.

• extradata : depending on mode parameter, extradata will contain one of the following: 1) a SVM object returned by e1071::tune.svm, containing both the best performing model and the description of the training process 2) a newly trained SVR model 3) the same object passed as tuned.model

See Also

svm, tune.svm, best.svm from e1071 package

Other phenoRegressors: phenoRegressor.BGLR(), phenoRegressor.RFR(), phenoRegressor.dummy(), phenoRegressor.rrBLUP(), phenoregressor.BGLR.multikinships()

Examples

## Not run: 
### WARNING ###
#The 'tuning' part of the example can take quite some time to run,
#depending on the computational power.

#using the GROAN.KI dataset, we regress on the dataset and predict the first ten phenotypes
phenos = GROAN.KI$yield
phenos[1:10] = NA

#--------- TUNE ---------
#tuning the SVR on a grid of hyperparameters
results.tune = phenoRegressor.SVR(
  phenotypes = phenos,
  genotypes = GROAN.KI$SNPs,
  covariances = NULL,
  extraCovariates = NULL,
  mode = 'tune',
  kernel = 'linear', cost = 10^(-3:+3) #SVR-specific parameters
)

#examining the predictions
plot(GROAN.KI$yield, results.tune$predictions,
     main = 'Mode = TUNING\nTrain set (black) and test set (red) regressions',
     xlab = 'Original phenotypes', ylab = 'Predicted phenotypes')
points(GROAN.KI$yield[1:10], results.tune$predictions[1:10], pch=16, col='red')

#printing correlations
test.set.correlation  = cor(GROAN.KI$yield[1:10], results.tune$predictions[1:10])
train.set.correlation = cor(GROAN.KI$yield[-(1:10)], results.tune$predictions[-(1:10)])
writeLines(paste(
  'test-set correlation :', test.set.correlation,
  '\ntrain-set correlation:', train.set.correlation
))

#--------- TRAIN ---------
#training the SVR, hyperparameters are given
results.train = phenoRegressor.SVR(
  phenotypes = phenos,
  genotypes = GROAN.KI$SNPs,
  covariances = NULL,
  extraCovariates = NULL,
  mode = 'train',
  kernel = 'linear', cost = 0.01 #SVR-specific parameters
)

#examining the predictions
plot(GROAN.KI$yield, results.train$predictions,
     main = 'Mode = TRAIN\nTrain set (black) and test set (red) regressions',
     xlab = 'Original phenotypes', ylab = 'Predicted phenotypes')
points(GROAN.KI$yield[1:10], results.train$predictions[1:10], pch=16, col='red')

#printing correlations
test.set.correlation  = cor(GROAN.KI$yield[1:10], results.train$predictions[1:10])
train.set.correlation = cor(GROAN.KI$yield[-(1:10)], results.train$predictions[-(1:10)])
writeLines(paste(
  'test-set correlation :', test.set.correlation,
  '\ntrain-set correlation:', train.set.correlation
))

#--------- RUN ---------
#we recover the trained model from previous run, predictions will be exactly the same
results.run = phenoRegressor.SVR(
  phenotypes = phenos,
  genotypes = GROAN.KI$SNPs,
  covariances = NULL,
  extraCovariates = NULL,
  mode = 'run',
  tuned.model = results.train$extradata
)

#examining the predictions
plot(GROAN.KI$yield, results.run$predictions,
     main = 'Mode = RUN\nTrain set (black) and test set (red) regressions',
     xlab = 'Original phenotypes', ylab = 'Predicted phenotypes')
points(GROAN.KI$yield[1:10], results.run$predictions[1:10], pch=16, col='red')

#printing correlations
test.set.correlation  = cor(GROAN.KI$yield[1:10], results.run$predictions[1:10])
train.set.correlation = cor(GROAN.KI$yield[-(1:10)], results.run$predictions[-(1:10)])
writeLines(paste(
  'test-set correlation :', test.set.correlation,
  '\ntrain-set correlation:', train.set.correlation
))

## End(Not run)

GROAN documentation built on Nov. 28, 2022, 5:07 p.m.