NEWS

hal9001 0.4.6 (2023-11-14)

Fixed predict method to address changes required by Matrix 1.6.2
Reduced default number of knots used for zero-order and first-order basis expansions as follows:
- base_num_knots_0 set to 20, used to be 200
- base_num_knots_1 set to 10, used to be 50

Fixed bug with prediction_bounds (a fit_hal argument in fit_control list), which would error when it was specified as a numeric vector. Also, added a check to assert this argument is correctly specified, and tests to ensure a numeric vector of bounds is provided.
Simplified fit_control list arguments in fit_hal. Users can still specify additional arguments to cv.glmnet and glmnet in this list.
Defined weights as a formal argument in fit_hal, opposed to an optional argument in fit_control, to facilitate specification and avoid confusion. This increases flexibility with SuperLearner wrapper SL.hal9001 as well; fit_control can now be customized with SL.hal9001.

As of September 2021:

Minor change to how binning is performed when num_knots = 1, ensuring that the minimal number of knots is chosen when num_knots = 1. This results in HAL agreeing with (main terms) glmnet when smoothness_orders = 1 and num_knots = 1.
Revised formula interface with enhanced capabilities, allowing specifciation of penalization factors, smoothness_orders, and the number of knots for each variable, for every single term separately using the new h function. It is possible to specify, e.g., h(X) + h(W) which will generate and concatenate the two basis function terms.

As of April 2021:

The default of fit_hal is now a first order smoothed HAL with binning.
Updated documentation for formula_hal, fit_hal and predict; and added fit_control and formula_control lists for arguments. Moved much of the text to details sections, and shortened the argument descriptions.
Updated summary to support higher-order HAL fit interpretations.
Added checks to fit_hal for missingness and dimensionality correspondence between X, Y, and X_unpenalized. These checks lead to quickly-produced errors, opposed to enumerating the basis list and then letting glmnet error on something trivial like this.
Modified formula interface in fit_hal, so formula is now provided directly to fit_hal and formula_hal is run within fit_hal. Due to these changes, it no longer made sense for formula_hal to accept data, so it now takes as input X. Also, the formula_fit_hal function was removed as it is no longer needed.
Support for the custom lasso procedure implemented in Rcpp has been discontinued. Accordingly, the "lassi" option and argument fit_type have been removed from fit_hal.
Re-added lambda.min.ratio as a fit_control argument to fit_hal. We've seen that not setting lambda.min.ratio in glmnet can lead to no lambda values that fit the data sufficiently well, so it seems appropriate to override the glmnet default.

As of February 2021:

Support higher order HAL via the new smoothness_orders argument
- smoothness_orders is a vector of length 1 or length ncol(X).
- If smoothness_orders is of length 1 then its values are recycled to form a vector of length ncol(X).
- Given such a vector of length ncol(X), the ith element gives the level of smoothness for the variable corresponding to the ith column in X.
Degree-dependant binning. Higher order terms are binned more coarsely; the num_knots argument is a vector up to max_degree controlling the degree-specific binning.
Adds formula_hal which allows a formula specification of a HAL model.

As of November 2020:

Allow support for Poisson family to glmnet().
Begins consideration of supporting arbitrary stats::family() objects to be passed through to calls to glmnet().
Simplifies output of fit_hal() by unifying the redundant hal_lasso and glmnet_lasso slots into the new lasso_fit slot.
Cleans up of methods throughout and improves documentation, reducing a few redundancies for cleaner/simpler code in summary.hal9001.
Adds link to DOI of the published Journal of Open Source Software paper in DESCRIPTION.

As of September 2020:

Adds a summary method for interpreting HAL regressions (https://github.com/tlverse/hal9001/pull/64).
Adds a software paper for publication in the Journal of Open Source Software (https://github.com/tlverse/hal9001/pull/71).

As of June 2020:

Address bugs/inconsistencies reported in the prediction method when trying to specify a value of lambda not included in initial fitting.
Addresses a bug arising from a silent failure in glmnet in which it ignores the argument lambda.min.ratio when family = "gaussian" is not set.
Adds a short software paper for submission to JOSS.
Minor documentation updates.

As of March 2020