,
Kaisa Joki [aut],
Anni Halkola [aut]| Title: | Optimal Subset Cardinality Regression (OSCAR) Models Using the L0-Pseudonorm |
|---|---|
| Description: | Optimal Subset Cardinality Regression (OSCAR) models offer regularized linear regression using the L0-pseudonorm, conventionally known as the number of non-zero coefficients. The package estimates an optimal subset of features using the L0-penalization via cross-validation, bootstrapping and visual diagnostics. Effective Fortran implementations are offered along the package for finding optima for the DC-decomposition, which is used for transforming the discrete L0-regularized optimization problem into a continuous non-convex optimization task. These optimization modules include DBDC ('Double Bundle method for nonsmooth DC optimization' as described in Joki et al. (2018) <doi:10.1137/16M1115733>) and LMBM ('Limited Memory Bundle Method for large-scale nonsmooth optimization' as in Haarala et al. (2004) <doi:10.1080/10556780410001689225>). The OSCAR models are comprehensively exemplified in Halkola et al. (2023) <doi:10.1371/journal.pcbi.1010333>). Multiple regression model families are supported: Cox, logistic, and Gaussian. |
| Authors: | Teemu Daniel Laajala [aut, cre]
,
Kaisa Joki [aut],
Anni Halkola [aut] |
| Maintainer: | Teemu Daniel Laajala <[email protected]> |
| License: | GPL-3 |
| Version: | 1.2.1 |
| Built: | 2024-10-22 04:16:11 UTC |
| Source: | https://github.com/syksy/oscar |
OSCAR models utilize the L0-pseudonorm to select an optimal subset of features that generalizes linear regression models to a variety of families. Currently supported models include conventional Gaussian regression (family="mse" or family="gaussian"), Binomial/Logistic regression (family="logistic"), and Cox proportional hazards modeling (family="cox").
Halkola AS, Joki K, Mirtti T, Mäkelä MM, Aittokallio T, Laajala TD (2023) OSCAR: Optimal subset cardinality regression using the L0-pseudonorm with applications to prognostic modelling of prostate cancer. PLoS Comput Biol 19(3): e1010333. doi:10.1371/journal.pcbi.1010333
Extract coefficients of oscar-objects
Prediction based on oscar-objects
Plot oscar-coefficients as a function of k and override default plot generic
## S4 method for signature 'oscar' coef(object, k) ## S4 method for signature 'oscar' predict( object, k, type = c("response", "link", "nonzero", "coefficients", "label"), newdata = object@x ) ## S4 method for signature 'oscar' plot(x, y, k = 1:x@kmax, add = FALSE, intercept = FALSE, ...)## S4 method for signature 'oscar' coef(object, k) ## S4 method for signature 'oscar' predict( object, k, type = c("response", "link", "nonzero", "coefficients", "label"), newdata = object@x ) ## S4 method for signature 'oscar' plot(x, y, k = 1:x@kmax, add = FALSE, intercept = FALSE, ...)
object |
Fit oscar S4-object |
k |
Vector of cardinality 'k' values |
type |
Type of prediction; valid values are 'response', 'link', 'nonzero', 'coefficients', or 'label' |
newdata |
Data to predict on; if no alternate is supplied, the function uses the original 'x' data matrix used to fit object |
x |
Values on x-axis |
y |
Values on y-axis |
add |
Should the plot be added on top of an existing plot (if FALSE, create a new graphics device), Default: FALSE |
intercept |
Should model intercept be plotted, Default: FALSE |
... |
Additional parameters passed on to the points-function drawing lines as a function of cardinality |
Vector of model coefficient values at given cardinality 'k'
A vector of coefficient predictions at the specificied cardinality 'k' with a format depending on the supplied 'type' parameter
Override default plot function with no return but instead tailor suitable graphics plotting
Return total cost of model fit based on provided kit/variable costs vector
Return total cost of model fit based on provided kit/variable costs vector
cost(object, k) ## S4 method for signature 'oscar' cost(object, k)cost(object, k) ## S4 method for signature 'oscar' cost(object, k)
object |
Fit oscar S4-object |
k |
Cardinality 'k' to compute total feature cost at |
Numeric value of total feature/kit cost at cardinality 'k'
Numeric value of total feature/kit cost at cardinality 'k'
An example data set from mCRPC patients in TYKS, along with cost vector / kit structure from HUSLAB
data(ex)data(ex)
data(ex)data(ex)
Return named vector of feature indices with a given k that are non-zero
Return named vector of feature indices with a given k that are non-zero
feat(object, k) ## S4 method for signature 'oscar' feat(object, k)feat(object, k) ## S4 method for signature 'oscar' feat(object, k)
object |
Fit oscar S4-object |
k |
Cardinality 'k' to extract non-zero features at |
Vector of feature indices at cardinality 'k'
Vector of feature indices at cardinality 'k'
Return named vector of indices for kits with a given k that are non-zero
Return named vector of indices for kits with a given k that are non-zero
kits(object, k) ## S4 method for signature 'oscar' kits(object, k)kits(object, k) ## S4 method for signature 'oscar' kits(object, k)
object |
Fit oscar S4-object |
k |
Cardinality 'k' to extract kit indices at |
Vector of kit indices at cardinality 'k'
Vector of kit indices at cardinality 'k'
This function fits an OSCAR model object to the provided training data with the desired model family.
oscar( x, y, k, w, family = "cox", metric, solver = 1, verb = 1, print = 3, kmax, sanitize = TRUE, percentage = 1, in_selection = 1, storeX = TRUE, storeY = TRUE, control, ... )oscar( x, y, k, w, family = "cox", metric, solver = 1, verb = 1, print = 3, kmax, sanitize = TRUE, percentage = 1, in_selection = 1, storeX = TRUE, storeY = TRUE, control, ... )
x |
Data matrix 'x' |
y |
Response vector/two-column matrix 'y' (see: family); number of rows equal to nrow(x) |
k |
Integer (0/1) kit indicator matrix; number of columns equal to ncol(x), Default: Unit diagonal indicator matrix |
w |
Kit cost weight vector w of length nrow(k), Default: Equal cost for all variables |
family |
Model family, should be one of: 'cox', 'mse'/'gaussian', or 'logistic, Default: 'cox' |
metric |
Goodness metric, Default(s): Concordance index for Cox, MSE for Gaussian, and AUC for logistic regression |
solver |
Solver used in the optimization, should be 1/'DBDC' or 2/'LMBM', Default: 1. |
verb |
Level of verbosity in R, Default: 1 |
print |
Level of verbosity in Fortran (may not be visible on all terminals); should be an integer between range, range, Default: 3 |
kmax |
Maximum k step tested, by default all k are tested from k to maximum dimensionality, Default: ncol(x) |
sanitize |
Whether input column names should be cleaned of potentially problematic symbols, Default: TRUE |
percentage |
Percentage of possible starting points used within range [0,1], Default: 1 |
in_selection |
Which starting point selection strategy is used (1, 2 or 3), Default: 1 |
storeX |
If data matrix X should be saved in the model object; turning this off might would help with memory, Default: TRUE |
storeY |
If data response Y should be saved in the model object; turning this off might would help with memory, Default: TRUE |
control |
Tuning parameters for the optimizers, see function oscar.control(), Default: see ?oscar.control |
... |
Additional parameters |
OSCAR utilizes the L0-pseudonorm, also known as the best subset selection, and makes use of a DC-formulation of the discrete feature selection task into a continuous one. Then an appropriate optimization algorithm is utilized to find optima at different cardinalities (k). The S4 model objects 'oscar' can then be passed on to various down-stream functions, such as oscar.pareto, oscar.cv, and oscar.bs, along with their supporting visualization functions.
Fitted oscar-object
oscar.cv oscar.bs oscar.pareto oscar.visu oscar.cv.visu oscar.bs.visu oscar.pareto.visu oscar.binplot
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit }
Create a sparse matrix with binary indicator 1 indicating that a coefficient was non-zero, and value 0 (or . in sparse matrix) indicating that a coefficient was zero (i.e. feature not included)
oscar.binarize(fit, kmax = fit@kmax)oscar.binarize(fit, kmax = fit@kmax)
fit |
Fit oscar-model object |
kmax |
Create matrix until kmax-value; by default same as for fit object, but for high dimensional tasks one may wish to reduce this |
The matrix consists of TRUE/FALSE values, and is very similar to the oscar.sparsify, where the function provides estimate values in a sparse matrix format.
A binary logical indicator matrix of variables (rows) as a function of cardinality k (columns), where elements are binary indicators for 1 as non-zero and 0 as zero.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.binarize(fit, kmax=5) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.binarize(fit, kmax=5) }
This visualization function makes use of the sparsified beta-coefficient matrix form as a function of cardinality. Optionally, user may showcase cross-validation performance alongside at the same cardinality values.
oscar.binplot( fit, cv, kmax, collines = TRUE, rowlines = TRUE, cex.axis = 0.6, heights = c(0.2, 0.8), ... )oscar.binplot( fit, cv, kmax, collines = TRUE, rowlines = TRUE, cex.axis = 0.6, heights = c(0.2, 0.8), ... )
fit |
Fitted oscar S4-class object |
cv |
Matrix produced by oscar.cv; rows are cv-folds, cols are k-values |
kmax |
Maximum cardinality 'k' |
collines |
Should vertical lines be drawn to bottom part |
rowlines |
Should horizontal lines be drawn to highlight variables |
cex.axis |
Axis magnification |
heights |
Paneling proportions as a numeric vector of length 2 |
... |
Additional parameters passed on to hamlet::hmap |
This is a plotting function that does not return anything, but instead draws on a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) oscar.binplot(fit=fit, cv=fit_cv) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) oscar.binplot(fit=fit, cv=fit_cv) }
This model bootstraps the fitting of a given oscar object (re-fits the model for data that is equal in size but sampled with replacement). The output objects give insight into robustness of the oscar-coefficient path, as well as relative importance of model objects.
oscar.bs(fit, bootstrap = 100, seed = NULL, verb = 0, ...)oscar.bs(fit, bootstrap = 100, seed = NULL, verb = 0, ...)
fit |
oscar-model object |
bootstrap |
Number of bootstrapped datasets, Default: 100 |
seed |
Random seed for reproducibility with NULL indicating that it is not set, Default: NULL |
verb |
Level of verbosity with higher integer giving more information, Default: 0 |
... |
Additional parameters passed to oscar-function |
The function provides a fail-safe try-catch in an event of non-convergence of the model fitting procedure. This may occur for example if a bootstrapped data matrix has a column consist of a single value only over all observations.
3-dimensional array with dimensions corresponding to k-steps, beta coefficients, and bootstrap runs
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.cv(fit, bootstrap = 20, seed = 123) fit_bs }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.cv(fit, bootstrap = 20, seed = 123) fit_bs }
This function plots as barplots as a function of k-cardinality in what proporties certain coefficients were chosen as non-zero over the bootstrap runs.
oscar.bs.boxplot(bs, ...)oscar.bs.boxplot(bs, ...)
bs |
Bootstrapped 3-dimensional array for an oscar object as produced by oscar.bs |
... |
Additional parameters passed on to barplot |
This is a plotting function that does not return anything, but instead draws on a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.boxplot(fit_bs) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.boxplot(fit_bs) }
The function reformats bootstrapped runs to a single long data.frame, where all bootstrapped runs are covered along with the choices for the variables at each cardinality 'k'.
oscar.bs.k(bs)oscar.bs.k(bs)
bs |
Bootstrapped list from oscar.bs |
Reformatted data.frame
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) ll <- oscar.bs.k(fit_bs) head(ll) tail(ll) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) ll <- oscar.bs.k(fit_bs) head(ll) tail(ll) }
This function neatly plots a colourized proportion of variables chosen as a function of cardinalities over a multitude of bootstrap runs. This helps model diagnostics in assesssing variable importance.
oscar.bs.plot( fit, bs, kmax, cex.axis = 0.6, palet = colorRampPalette(c("orange", "red", "black", "blue", "cyan"))(dim(bs)[3]), nbins = dim(bs)[3], Colv = NA, Rowv = NA, ... )oscar.bs.plot( fit, bs, kmax, cex.axis = 0.6, palet = colorRampPalette(c("orange", "red", "black", "blue", "cyan"))(dim(bs)[3]), nbins = dim(bs)[3], Colv = NA, Rowv = NA, ... )
fit |
Fitted oscar S4-class object |
bs |
Bootstrapped 3-dimensional array for an oscar object as produced by oscar.bs |
kmax |
Maximum cardinality 'k' |
cex.axis |
Axis magnification |
palet |
Colour palette |
nbins |
Number of bins (typically ought to be same as number of colours in the palette) |
Colv |
Column re-ordering indices or a readily built dendrogram |
Rowv |
Row re-ordering indices or a readily built dendrogram |
... |
Additional parameters passed on to the hamlet::hmap function |
Further heatmap parameters available from ?hmap
This is a plotting function that does not return anything, but instead draws on a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.plot(fit, fit_bs) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.plot(fit, fit_bs) }
This function visualizes bootstrapped model coefficients over multiple bootstrap runs as lines in a graph
oscar.bs.visu(bs, intercept = FALSE, add = FALSE)oscar.bs.visu(bs, intercept = FALSE, add = FALSE)
bs |
Bootstrapped 3-dimensional array for an oscar object as produced by oscar.bs |
intercept |
Whether model intercept should be plotted also as a coefficient, Default: FALSE |
add |
Should plot be added on top of an existing plot device |
This is a plotting function that does not return anything, but instead draws on an existing or a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.visu(fit_bs) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_bs <- oscar.bs(fit, bootstrap = 20, seed = 123) oscar.bs.visu(fit_bs) }
Fine-tuning the parameters available for the DBDC and LMBM optimizers. See oscar documentation for the optimization algorithms for further details.
oscar.control( x, family, start = 2, in_mrounds = 5000, in_mit = 5000, in_mrounds_esc = 5000, in_b1, in_b2 = 3, in_b, in_m = 0.01, in_m_clarke = 0.01, in_c = 0.1, in_r_dec, in_r_inc = 10^5, in_eps1 = 5 * 10^(-5), in_eps, in_crit_tol = 10^(-5), na = 4, mcu = 7, mcinit = 7, tolf = 10^(-5), tolf2 = 10^4, tolg = 10^(-5), tolg2 = tolg, eta = 0.5, epsL = 0.125 )oscar.control( x, family, start = 2, in_mrounds = 5000, in_mit = 5000, in_mrounds_esc = 5000, in_b1, in_b2 = 3, in_b, in_m = 0.01, in_m_clarke = 0.01, in_c = 0.1, in_r_dec, in_r_inc = 10^5, in_eps1 = 5 * 10^(-5), in_eps, in_crit_tol = 10^(-5), na = 4, mcu = 7, mcinit = 7, tolf = 10^(-5), tolf2 = 10^4, tolg = 10^(-5), tolg2 = tolg, eta = 0.5, epsL = 0.125 )
x |
Input data matrix 'x'; will be used for calculating various control parameter defaults. |
family |
Model family; should be one of 'cox', 'logistic', or 'gaussian'/'mse' |
start |
Starting point generation method, see vignettes for details; should be an integer between range,range, Default: 2 |
in_mrounds |
DBDC: The maximum number of rounds in one main iteration, Default: 5000 |
in_mit |
DBDC: The maximum number of main iterations, Default: 5000 |
in_mrounds_esc |
DBDC: The maximum number of rounds in escape procedure, Default: 5000 |
in_b1 |
DBDC: The size of bundle B1, Default: min(n_feat+5,1000) |
in_b2 |
DBDC: The size of bundle B2, Default: 3 |
in_b |
DBDC: Bundle B in escape procedure, Default: 2*n_feat |
in_m |
DBDC: The descent parameter in main iteration, Default: 0.01 |
in_m_clarke |
DBDC: The descent parameter in escape procedure, Default: 0.01 |
in_c |
DBDC: The extra decrease parameter in main iteration, Default: 0.1 |
in_r_dec |
DBDC: The decrease parameter in main iteration, Default: 0.75, 0.99, or larger depending on n_obs (thresholds 10, 300, and above) |
in_r_inc |
DBDC: The increase parameter in main iteration, Default: 10^5 |
in_eps1 |
DBDC: The enlargement parameter, Default: 5*10^(-5) |
in_eps |
DBDC: The stopping tolerance (proximity measure), Default: 10^(-6) if number of features is <= 50, otherwise 10^(-5) |
in_crit_tol |
DBDC: The stopping tolerance (criticality tolerance), Default: 10^(-5) |
na |
LMBM: Size of the bundle, Default: 4 |
mcu |
LMBM: Upper limit for maximum number of stored corrections, Default: 7 |
mcinit |
LMBM: Initial maximum number of stored corrections, Default: 7 |
tolf |
LMBM: Tolerance for change of function values, Default: 10^(-5) |
tolf2 |
LMBM: Second tolerance for change of function values, Default: 10^4 |
tolg |
LMBM: Tolerance for the first termination criterion, Default: 10^(-5) |
tolg2 |
LMBM: Tolerance for the second termination criterion, Default: same as 'tolg' |
eta |
LMBM: Distance measure parameter (>0), Default: 0.5 |
epsL |
LMBM: Line search parameter (0 < epsL < 0.25), Default: 0.125 |
This function sanity checks and provides reasonable DBDC ('Double Bundle method for nonsmooth DC optimization' as described in Joki et al. (2018) <doi:10.1137/16M1115733>) and LMBM ('Limited Memory Bundle Method for large-scale nonsmooth optimization' as presented in Haarala et al. (2004) <doi:10.1080/10556780410001689225>) optimization tuning parameters. User may override custom values, though sanity checks will prevent unreasonable values and replace them. The returned list of parameters can be provided for the 'control' parameter when fitting oscar-objects.
A list of sanity checked parameter values for the OSCAR optimizers.
if(interactive()){ oscar.control() # Return a list of default parameters }if(interactive()){ oscar.control() # Return a list of default parameters }
If at least one measurement from a kit is included in the model, the kit cost is added.
oscar.cost.after(object)oscar.cost.after(object)
object |
Fit oscar S4-object |
A vector for numeric values of total kit costs at different cardinalities.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.cost.after(fit) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.cost.after(fit) }
Create a cross-validation matrix with the chosen goodness metric with n-folds. Based on the goodness metric, one ought to pick optimal cardinality (parameter 'k').
oscar.cv( fit, fold = 10, seed = NULL, strata = rep(1, times = nrow(fit@x)), verb = 0, ... )oscar.cv( fit, fold = 10, seed = NULL, strata = rep(1, times = nrow(fit@x)), verb = 0, ... )
fit |
oscar-model object |
fold |
Number of cross-validation folds, Default: 10 |
seed |
Random seed for reproducibility with NULL indicating that it is not set, Default: NULL |
strata |
Should stratified cross-validation be used; separate values indicate balanced strata. Default: Unit vector, which will treat all observations equally. |
verb |
Level of verbosity with higher integer giving more information, Default: 0 |
... |
Additional parameters passed to oscar-function |
A k-fold cross-validation is run by mimicking the parameters contained in the original oscar S4-object. This requires the original data at slots @x and @y.
A matrix with goodness of fit over folds and k-values
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold=10, seed=123) fit_cv }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold=10, seed=123) fit_cv }
This function plots the model performance as a function of cardinality for k-fold cross-validation. Performance metric depends on user choice and model family (i.e. lower MSE is good, higher C-index is good).
oscar.cv.visu( cv, add = FALSE, main = "OSCAR cross-validation", xlab = "Cardinality 'k'", ylab = "CV performance", ... )oscar.cv.visu( cv, add = FALSE, main = "OSCAR cross-validation", xlab = "Cardinality 'k'", ylab = "CV performance", ... )
cv |
Matrix produced by oscar.cv; rows are cv-folds, cols are k-values |
add |
Should plot be added on top of an existing plot device |
main |
Main title |
xlab |
X-axis label |
ylab |
Y-axis label |
... |
Additional parameters passed on top the CV points |
This is a plotting function that does not return anything, but instead draws on an existing or a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) oscar.cv.visu(fit_cv) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) oscar.cv.visu(fit_cv) }
This function retrieves the set of pareto optimal points for an oscar model fit in n-proportional time as cardinality axis is readily sorted. It is advisable to optimize model generalization (via cross-validation) rather than mere goodness-of-fit.
oscar.pareto(fit, cv, xval = "cost", weak = FALSE, summarize = mean)oscar.pareto(fit, cv, xval = "cost", weak = FALSE, summarize = mean)
fit |
Fit oscar S4-object |
cv |
A cross-validation matrix as produced by oscar.cv; if CV is not provided, then goodness-of-fit from fit object itself is used rather than cross-validation generalization metric |
xval |
The x-axis to construct pareto front based on; by default 'cost' vector for features/kits, can also be 'cardinality'/'k' |
weak |
If weak pareto-optimality is allowed; by default FALSE. |
summarize |
Function that summarizes over cross-validation folds; by default, this is the mean over the k-folds. |
A data.frame containing points and indices at which pareto optimal points exist
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold=10) oscar.pareto(fit, cv=fit_cv) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold=10) oscar.pareto(fit, cv=fit_cv) }
Visualization function for showing the pareto front for cardinality 'k' and model goodness metric, either from goodness-of-fit or from cross-validation
oscar.pareto.visu( fit, cv, xval = "cost", weak = FALSE, summarize = mean, add = FALSE, ... )oscar.pareto.visu( fit, cv, xval = "cost", weak = FALSE, summarize = mean, add = FALSE, ... )
fit |
Fit oscar S4-object |
cv |
A cross-validation matrix as produced by oscar.cv; if CV is not provided, then goodness-of-fit from fit object itself is used rather than cross-validation generalization metric |
xval |
The x-axis to construct pareto front based on; by default 'cost' vector for features/kits, can also be 'cardinality'/'k' |
weak |
If weak pareto-optimality is allowed; by default FALSE. |
summarize |
Function that summarizes over cross-validation folds; by default, this is the mean over the k-folds. |
add |
If the fit should be added on top of an existing plot; in that case leaving out labels etc. By default new plot is called. |
... |
Additional parameters provided for the plotting functions |
This is a plotting function that does not return anything, but instead draws on an existing or a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) opar <- par(mfrow=c(1,2)) oscar.pareto.visu(fit=fit) # Model goodness-of-fit oscar.pareto.visu(fit=fit, cv=fit_cv) # Model cross-validation performance par(opar) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') fit_cv <- oscar.cv(fit, fold = 10, seed = 123) opar <- par(mfrow=c(1,2)) oscar.pareto.visu(fit=fit) # Model goodness-of-fit oscar.pareto.visu(fit=fit, cv=fit_cv) # Model cross-validation performance par(opar) }
Variable estimates (rows) as a function of cardinality (k, columns). Since a model can drop out variables in favor of two better ones as k increases, this sparse representation helps visualize which variables are included at what cardinality.
oscar.sparsify(fit, kmax = fit@kmax)oscar.sparsify(fit, kmax = fit@kmax)
fit |
oscar-model object |
kmax |
Create matrix until kmax-value; by default same as for fit object, but for high dimensional tasks one may wish to reduce this |
Uses sparseMatrix-class from Matrix-package
A sparse matrix of variables (rows) as a function of cardinality k (columns), where elements are the beta estimates.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.sparsify(fit, kmax=5) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.sparsify(fit, kmax=5) }
Plot oscar S4-object goodness-of-fit, kit costs, and similar performance metrics.
oscar.visu( fit, y = c("target", "cost", "goodness", "cv", "AIC"), cols = c("red", "blue"), legend = "top", mtexts = TRUE, add = FALSE, main = "" )oscar.visu( fit, y = c("target", "cost", "goodness", "cv", "AIC"), cols = c("red", "blue"), legend = "top", mtexts = TRUE, add = FALSE, main = "" )
fit |
Fitted oscar S4-class object |
y |
Plotted y-axes supporting two simultaneous axes with different scales, Default: c("target", "cost", "goodness", "cv") |
cols |
Colours for drawn lines, Default: c("red", "blue") |
legend |
Location of legend or omission of legend with NA, Default: 'top' |
mtexts |
Outer margin texts |
add |
Should plot be added into an existing frame / plot |
main |
Main title |
This is a plotting function that does not return anything, but instead draws on an existing or a new graphics device.
if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.visu(fit, y=c("target", "cost")) }if(interactive()){ data(ex) fit <- oscar(x=ex_X, y=ex_Y, k=ex_K, w=ex_c, family='cox') oscar.visu(fit, y=c("target", "cost")) }
Showing oscar-objects
## S4 method for signature 'oscar' show(object)## S4 method for signature 'oscar' show(object)
object |
Fit oscar S4-object |
Outputs raw text describing key characteristics of the oscar-object