mbonakda / ranger

A Fast Implementation of Random Forests

ranger: A Fast Implementation of Random Forests

• Marvin N. Wright, [e-mail]
• Extended by Matt Bonakdarpour

Introduction

This fork of the ranger package contains implementations of the following two methods:

• Random Forest Reshaping (Bonakdarpour et al., 2018)
• Discrete Choice Random Forests

Reshaping

The reshaping implementation allows the user to input a character vector corresponding to the desired reshaped predictor variables. The reshaping algorithm will enforce the constraint such that the predictions are monotonically increasing with respect to those input variables.

The current reshaping functionality is only implemented for regression and probability trees. In the master branch, only the "over-constrained estimator" (Bonakdarpour et al, 2018) is implemented to minimize dependencies on external packages. The reshape-only branch also contains the "exact estimator" but requires updates to the Makevars file based on your local mosek installation path.

Here is example usage for fitting a shape-constrained random forest with respect to an input variabled called bmi:

rf                       <- ranger(dependent.variable.name = 'y',
                                   data                    = input.data.df,
                                   write.forest            = TRUE,
                                   num.trees               = 100,
                                   sc.variable.names       = c('bmi')
                                   )

Discrete Choice Random Forests

The discrete choice random forest implementation allows the user to specify a speedy input parameter which implements an approximate split-finding algorithm for efficiency. The input dataframe is expected to be in "long" format as defined by the mlogit R package.

The implementation expects a column labeled agentID which uniquely identifies the agent making the choice in the corresponding row. Each row contains information about a choice event -- predictor variables for the agent and a single item. Multiple rows will correspond to an agent. The dependent variable is assumed to be an integer (0 or 1) depending on whether or not the agent chose the corresponding item in that row. The remaining columns are assumed to be predictor variables.

Example usage:

rf                      <- ranger(dependent.variable.name = 'choice',
                                  data                = input.data.df,
                                  write.forest        = TRUE,
                                  num.trees           = 100,
                                  discrete.choice     = TRUE,
                                  speedy              = TRUE
                                 )

References

• Wright, M. N. & Ziegler, A. (2017). ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R. Journal of Statistical Software 77:1-17. http://dx.doi.org/10.18637/jss.v077.i01.
• Schmid, M., Wright, M. N. & Ziegler, A. (2016). On the use of Harrell’s C for clinical risk prediction via random survival forests. Expert Systems with Applications 63:450-459. http://dx.doi.org/10.1016/j.eswa.2016.07.018.
• Wright, M. N., Dankowski, T. & Ziegler, A. (2017). Unbiased split variable selection for random survival forests using maximally selected rank statistics. Statistics in Medicine. http://dx.doi.org/10.1002/sim.7212.
• Breiman, L. (2001). Random forests. Machine learning 45:5-32.
• Ishwaran, H., Kogalur, U. B., Blackstone, E. H., & Lauer, M. S. (2008). Random survival forests. The Annals of Applied Statistics 2:841-860.
• Malley, J. D., Kruppa, J., Dasgupta, A., Malley, K. G., & Ziegler, A. (2012). Probability machines: consistent probability estimation using nonparametric learning machines. Methods of Information in Medicine 51:74-81.
• Bonakdarpour, M., Chatterjee, S., Foygel Barber, R., Lafferty, J. (2018). ICML