GameRank - An R package for Feature Selection

The GameRank package is an R package that implements functions for variable and feature selection to create predictive models. This includes a number of feature selection algorithms, mostly combinatorial search strategies, which are:

• Forward selection,
• Backward selection,
• Bidirectional search, and
• Plus-L, Minus-R search together with
• Random search, which draws random combinations across the feature space, and
• GameRank a novel feature selection algorithm, that employs a maximum likelihood ranking model to learn the optimal ranking of features. The following figure illustrates the algorithm:

In addition, it provides methods for feature construction, that is

• functions to screen variable properties,
• evaluate simple variable transformations,
• Box-Cox transformations for regression and binomial models, as well as
• methods for the detection of features with multi-modal distributions.

GameRank allows for direct optimization of measures related to calibration and discrimination of models. This approach uses a training:validation:test split-set approach. However, for small sample selection projects, functions for bias-corrections via cross-validation or bootstrapping are included.

Parts of this code have been used to craft a model for chemo-tolerability prediction. [https://ascopubs.org/doi/full/10.1200/CCI.21.00121]

The package can be installed directly from GitHub via

devtools::install_github("Genentech/GameRank")

Here are some few lines of example code to use the package to screen features:

library( dplyr )
library( ggplot2 )
vck <- check_variables( toy_data, response_var, list_variables )
vck %>% 
   ggplot(aes(x=entropy, y=mutual_information) ) +
   geom_point()

and do a forward selection:

sel <- forward( toy_data, response_var, list_variables 
                 fn_train_binomial,  # Train a logistic regression model
                 fn_eval_binomial_auroc,  # Evaluate the Area under ROC curve
                 m=4L,   # Search for combination of 4 features
                 splits=2L, # Evaluate on 2 parallel split sets
                 maximize=TRUE # Maximize the Area under ROC curve
                 )
sel$variable_selections

Have fun!