This is a MATLAB wrapper for doing classification using the decision forest framework [1] available at http://research.microsoft.com/en-us/projects/decisionforests/ , hereinafter referred to as Sherwood.

Training and classification is parallelized.

Getting started

1. MATLAB and a c++ compiler is needed.

2. Due to the license of Sherwood, you will have to download the source code and place it inside /Sherwood/.

3. Setup MATLAB using "mex -setup".

4. An example is available in example.m

All files are compiled automatically

The code has been tested on

• MATLAB 2013a with GCC 4.8 on Ubuntu 13.10.
• MATLAB 2013a with Visual Studio 2013 on Windows 7.

Limitations

If you are using a c++ compiler which does not support OpenMP http://openmp.org/wp/openmp-compilers/, you need to turn off multi-threaded training by setting "use_openmp = false" in sherwood_train.m.

OpenMP is unfortunately not supported by the c++ compiler recommended by Mathworks for Windows: Microsoft Windows SDK 7.1. It is however supported by Visual Studio.

Alternatives

There are atleast two popular implementation of random forest available for MATLAB [2,3].

The main difference with Sherwood are

1. No bagging, so there will be no out-of-bag error et cetera.
2. The probabilities in the leafs are stored as histograms resulting in more accurate probability estimates when averaging over many trees.
3. Entropy is used as split critera for each node.

References

1. Decision Forests: A Unified Framework for Classification, Regression, Density Estimation, Manifold Learning and Semi-Supervised Learning
   Foundations and Trends® in Computer Graphics and Vision, 2012.
   Antonio Criminisi, Jamie Shotton, and Ender Konukoglu.

2. Randomforest-matlab
   A Matlab wrapper of a R wrapper of Leo Breiman's fortran code.

3. Tree bagger
   Matlab's own random forest variation.