The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.

tramnet 0.1-0

New features

• tramnet() is now an S3 generic. The existing interface, taking an unpenalized tram model and a design matrix, is available as tramnet.default() and remains unchanged.
• New formula interface tramnet.formula(), the penalized analogue of tram::tram(), together with the convenience wrappers Coxphnet(), Colrnet(), Lmnet(), BoxCoxnet(), Lehmannnet() and Survregnet(). The design matrix is set up and scaled internally, and the scaling of the training data is re-used for newdata in all methods and in cvl_tramnet().
• coef() gains argument scaled. For models fitted via the formula interface, coefficients are reported on the original scale of the covariates by default.
• New model.matrix() method.
• mbo_recommended() additionally accepts a fitted "tramnet" object.
• Case weights and offsets are now taken into account when fitting.

Changes

• Model-based optimization now uses package ‘mlr3mbo’ instead of the retired package ‘mlrMBO’. Function names and arguments of mbo_tramnet(), elnet_obj(), lasso_obj(), ridge_obj() and mbo_recommended() are unchanged, but the default of opt_crit is now "ei", mbo_tramnet() returns a list and the objective functions are of class bbotk::ObjectiveRFun.
• Tests use ‘testthat’.
• Removed unused dependencies (‘mlbench’, ‘utils’) and the unused internal function .plot_cvl(). The package description no longer mentions models for binary and ordered categorical responses, which cannot be fitted.

Bug fixes

• logLik(object, newdata) now returns the out-of-sample log-likelihood without subtracting the penalty of the training fit. Consequently, cvl_tramnet() and mbo_tramnet() tune lambda and alpha based on the unpenalized out-of-sample log-likelihood (log score), which previously favoured small values of lambda.
• The quadratic part of the penalty was erroneously scaled by lambda^2 instead of lambda in the objective function. The penalty is now lambda * (0.5 * (1 - alpha) * ||beta||_2^2 + alpha * ||beta||_1), as documented and as in ‘glmnet’, and coincides with the penalty subtracted in logLik(). Fits with alpha < 1 and lambda different from 0 and 1 change.
• predict(), simulate(), plot(), residuals() and estfun() now use the correct sign of the shift term for models with negative = TRUE, such as Lm(), BoxCox(), Lehmann() and Survreg().
• plot_path() no longer warns when restoring graphical parameters.
• predict(), simulate() and plot() work with non-syntactic column names of the design matrix (e.g., interactions or log(x)).
• prof_lambda(), prof_alpha() and cvl_tramnet(..., fit_opt = TRUE) refit models directly instead of re-evaluating the original call, which failed when the baseline model or the design matrix were not visible from the global environment.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.