Skip to content

InfiniteBoost: building infinite ensembles with gradient descent

Notifications You must be signed in to change notification settings

arogozhnikov/infiniteboost

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

51 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

InfiniteBoost

Code for a paper
InfiniteBoost: building infinite ensembles with gradient descent (arXiv:1706.01109).
A. Rogozhnikov, T. Likhomanenko

Description

InfiniteBoost is an approach to building ensembles which combines best sides of random forest and gradient boosting.

Trees in the ensemble encounter mistakes done by previous trees (as in gradient boosting), but due to modified scheme of encountering contributions the ensemble converges to the limit, thus avoiding overfitting (just as random forest).

Left: InfiniteBoost with automated search of capacity vs gradient boosting with different learning rates (shrinkages), right: random forest vs InfiniteBoost with small capacities.

More plots of comparison in research notebooks and in research/plots directory.

Reproducing research

Research is performed in jupyter notebooks (if you're not familiar, read why Jupyter notebooks are awesome).

You can use the docker image arogozhnikov/pmle:0.01 from docker hub. Dockerfile is stored in this repository (ubuntu 16 + basic sklearn stuff).

To run the environment (sudo is needed on Linux):

sudo docker run -it --rm -v /YourMountedDirectory:/notebooks -p 8890:8890 arogozhnikov/pmle:0.01

(and open localhost:8890 in your browser).

InfiniteBoost package

Self-written minimalistic implementation of trees as used for experiments against boosting.

Specific implementation was used to compare with random forest and based on the trees from scikit-learn package.

Code written in python 2 (expected to work with python 3, but not tested), some critical functions in fortran, so you need gfortran + openmp installed before installing the package (or simply use docker image).

pip install numpy
pip install .
# testing (optional)
cd tests && nosetests .

You can use implementation of trees from the package for your experiments, in this case please cite InfiniteBoost paper.

About

InfiniteBoost: building infinite ensembles with gradient descent

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published