Warning: This version includes several breaking changes.
preference_order()Now works with any combination of categorical and numeric responses and predictors. Previously, only numeric responses were considered valid.
Accepts a character vector with multiple response variables and returns a named list of data frames in such cases.
All functions used as input for the argument f have
been rewritten, with extended coverage of cases. These functions have
also been consistently renamed following these rules:
r2 for R-squared,
auc for area under the curve (for binomial responses), and
v for Cramer’s V (for categorical responses).spearman,
pearson, and v for direct association;
glm for GLMs; gam for GAMs; rf
for Random Forest models; and rpart for Recursive Partition
Trees.gaussian for numeric
responses, binomial for binomial responses, and
poisson for integer counts.poly2 for GLMs with second-degree
polynomials.When f = NULL, the function f_auto()
determines an appropriate default adapted to the types of the response
and predictors.
Now issues a warning if predictors show a suspiciously high
association with the response. The sensitivity of this test is
controlled by the new argument warn_limit.
Parallelization setup is now managed via
future::plan(), and a progress bar is available through
progressr::handlers().
collinear()Now works with any combination of categorical and numeric responses and predictors. Previously, only numeric responses were valid. Categorical predictors are excluded from VIF analysis but are returned in the output if they pass the pairwise correlation test.
Accepts a character vector with multiple response variables and returns a named list of data frames in such cases.
The preference order is now computed internally if
preference_order = NULL (default). Therefore, all relevant
arguments of the function preference_order() have been
added to collinear() with the prefix
“preference_”.
Parallelization setup is now managed via
future::plan(), with a progress bar provided by
progressr::handlers(). This setup is leveraged by
preference_order() and cor_select().
Target encoding can be disabled by setting the
encoding_method argument to NULL.
VIF filtering can be disabled by setting max_vif to
NULL.
Pairwise correlation filtering can be disabled by setting
max_cor to NULL.
cor_select()A new robust forward selection algorithm ensures that the most
important predictors are retained after multicollinearity filtering when
preference_order is used.
Target encoding, along with the response and
encoding_method arguments, has been removed from this
function. This change also applies to cor_df().
The function now calls validate_data_cor() to ensure
that the data is suitable for pairwise correlation multicollinearity
filtering.
Parallelization setup is now managed via
future::plan(), with a progress bar provided by
progressr::handlers(). This setup is used by
cor_numeric_vs_categorical() and
cor_categorical_vs_categorical() to speed up pairwise
correlation computation.
cor_df()cor_numeric_vs_categorical()
and cor_categorical_vs_categorical() from triggering
properly.vif_select()A new robust forward selection algorithm better preserves
predictors with higher preference when preference_order is
used.
Target encoding, along with the response and
encoding_method arguments, has been removed. As a result,
this function now only works with numeric predictors. This change also
applies to vif_df().
The new function validate_data_vif() is called to
ensure the data is suitable for VIF-based multicollinearity filtering.
Attempting a VIF analysis in a data frame with more columns than rows
now returns an error.
target_encoding_lab() and Companion FunctionsCompletely rewritten for parallelization using
future::plan() and a progress bar via
progressr::handlers().
The default encoding method is now “loo” (leave-one-out), as it provides more useful results in most cases.
The functions target_encoding_mean(),
target_encoding_rank(), and
target_encoding_loo() have been simplified to the bare
minimum, with all redundant logic moved to
target_encoding_lab().
NA cases in the predictor to encode are now grouped under “NA”.
The “rnorm” method has been deprecated, and the function
target_encoding_rnorm() has been removed from the
package.
Added the function cor_clusters() to group
predictors using stats::hclust() based on their pairwise
correlation matrix.
Streamlined the package documentation using roxygen methods to inherit sections and parameters.
Removed dplyr as a dependency.
Added mgcv, rpart, and
ranger to Imports to support all f_xxx()
functions from the start.
All warnings in data validation functions have been converted to messages. These messages now indicate the function that generated them, aiding in debugging and ensuring that messages and warnings are printed in the correct order.
Hotfix of issue with solve(tol = 0) in systems with no large double support (noLD). This one wasn’t fun.
Added argument “smoothing” to target_encoding_mean()
function to implement original target encoding method.
Added alias f_rf_rsquared() to the function
f_rf_deviance().
Added column “vi_binary” to vi as a binary version of
“vi_mean”.
Added function auc_score() to compute the area under the
curve of predictions from binary models.
Added function case_weights() to compute case weights
when binary responses are unbalanced.
Added function f_rf_auc_balanced() to be used as input
for the f argument of preference_order() when
the response is binary and balanced.
Added function f_rf_auc_unbalanced() to be used as input
for the f argument of preference_order() when
the response is binary and unbalanced.
Added function f_gam_auc_balanced() to be used as input
for the f argument of preference_order() when
the response is binary and balanced.
Added function f_gam_auc_unbalanced() to be used as
input for the f argument of preference_order()
when the response is binary and unbalanced.
Added function f_logistic_auc_balanced() to be used as
input for the f argument of preference_order()
when the response is binary and balanced.
Added function f_logistic_auc_unbalanced() to be used as
input for the f argument of preference_order()
when the response is binary and unbalanced.
Fixed issue with perfect correlations in vif_df(). Now
perfect correlations are replaced with 0.99 (for correlation == 1) and
-0.99 (for correlation == -1) in the correlation matrix to avoid errors
in solve().
Added the example dataset toy, derived from
vi, but with known relationships between all variables.
Fixed issue in function cor_df() where many cases would be lost because the logic to remove diagonals was flawed, as all pairs with correlation == 1 were being removed.
Fixed issue in functions cor_select() and vif_select() where ignoring
predictors and using only df would lead to
empty selections.
This version fixes bugs in two functions: cor_select()
and cor_df()
cor_select()
When only one variable was left in the correlation matrix, the one column matrix became a vector with no colnames, which yielded an error. Now, to avoid this issue, drop = FALSE is used in the matrix subsetting.
The previous version started removing predictors on a backwards fashion, from the last predictor in preference order, moving up one by one to the top. Under the wrong circumstances (low number of predictors, low max_cor, and high correlation between first and second predictors in preference order) this configuration would lead to keep only the first predictor, even when having others comply with the max_cor restriction lower down in the preference order. The new version produces smaller subsets of predictors with a higher diversity.
cor_df()
Re-submission after minor CRAN comments.
Changes:
version number bumped up to 1.0.1
removed if(interactive()){} from all @examples.
removed plot() call from the @examples of the function target_encoding_lab() because it was messing up pkgdown’s build. This piece of code triggered the comment “Please always make sure to reset to user’s options()” by the reviewer, so this should solve the issue.
made sure that all examples run in less than 5 seconds.
fixed a bug in which all functions would include the response as a predictor when ‘predictors = NULL’ and ‘response’ was a valid column of the input data frame.
First functional version of the package submitted to CRAN.