Skip to content

Future work on FastRF

Jump to bottom
Fran Supek edited this page Sep 7, 2017 · 2 revisions

  • Systematically investigate the effects of the parameters for m_Kvalue (number of features considered in a node) and m_numFeatTree (number of features considered in a tree) on a larger number of datasets. Propose better default values.

  • Add support numeric class attributes (regression trees)

  • Change the one-vs-all binary split used for categorical attributes to a multifurcating split, as in Weka RF. This will improve the execution time on datasets with many multi-level categorical attributes.

  • Test the algorithm to calculate dropout feature importance - how it relates to standard feature importance measures?

  • Benchmark FastRF on many-core machines. Compare to other popular non-Java RF implementations.

  • Make FastRF into a Weka package, therefore easier to install into the Weka GUI.

  • Write a wrapper for R.

  • ... we'd be happy to hear your suggestions!

Clone this wiki locally