Please see the file CHANGES.md for the changelog of FastRGF.

1. Introduction

2. Installation

3. Examples

4. Hyperparameters Tuning

5. Known Issues

6. Contact

7. Copyright

8. References

This software package provides a multi-core implementation of a simplified Regularized Greedy Forest (RGF) described in [1]. Please cite the paper if you find the software useful.

RGF is a machine learning method for building decision forests that have been used to win some Kaggle competitions. In our experience it works better than gradient boosting on many relatively large datasets.

The implementation employs the following concepts described in the original paper [1]:

• tree node regularization;
• fully-corrective update;
• greedy node expansion with trade-off between leaf node splitting for current tree and root splitting for new tree.

However, various simplifications are made to accelerate the training speed. Therefore, unlike the original RGF program, this software does not reproduce the results in the paper.

The implementation of greedy tree node optimization employs second order Newton approximation for general loss functions. For logistic regression loss, which works especially well for many binary classification problems, this approach was considered in [2]; for general loss functions, second order approximation was considered in [3].

The software is written in C++11 and requires to be built from the sources with the g++, Clang or AppleClang compiler. Note that compilation only with g++-5, Clang 3.8, AppleClang 8.1 (or newer versions) is possible. Other compilers are unsupported and older versions produce corrupted files. In case you are compiling with the AppleClang compiler, you should install OpenMP separately.

Download the package and extract the content. Otherwise, you can use git

git clone https://github.com/RGF-team/rgf.git

The top directory of the extracted (or cloned) content is rgf and this software package is located into FastRGF subfolder. Below all the path expressions are relative to rgf/FastRGF.

The source files are located in the include and src directories.

The following executables will appear in the bin directory after you compile them by any method from the listed below.

• forest_train: train FastRGF and save the model;
• forest_predict: apply already trained model on test data.

You may use the option -h to show the command-line options of the each executable file.

On Windows compilation only with CMake and MinGW-w64 is supported because only this version of MinGW provides POSIX threads.

mkdir build
cd build
cmake .. -G "MinGW Makefiles"
mingw32-make
mingw32-make install

mkdir build
cd build
cmake ..
make
make install

Please go to the examples subdirectory.

• forest.ntrees: Controls the number of trees in the forest. Typical range is [100, 10000]. Default value is 500.
• forest.opt: Optimization method for training the forest. You can select rgf or epsilon-greedy. Default value is rgf.
• forest.stepsize: Controls the step size of epsilon-greedy boosting. Meant for being used with forest.opt=epsilon-greedy. Default value is 0.0.

• dtree.max_level: Controls the maximum tree depth. Default value is 6.
• dtree.max_nodes: Controls the maximum number of leaf nodes in best-first search. Default value is 50.
• dtree.new_tree_gain_ratio: Controls when to start a new tree. New tree is created when leaf nodes gain < this value * estimated gain of creating new tree. Default value is 1.0.
• dtree.min_sample: Controls the minimum number of training data points in each leaf node. Default value is 5.
• dtree.loss: You can select LS, MODLS or LOGISTIC loss function. Default value is LS. However, for binary classification task LOGISTIC often works better.
• dtree.lamL1: Controls the degree of L1 regularization. A large value induces sparsity. Typical range is [0.0, 1000.0]. Default value is 1.0.
• dtree.lamL2: Controls the degree of L2 regularization. The larger value is, the larger forest.ntrees you need to use: the resulting accuracy is often better with a longer training time. Use a relatively large value such as 1000.0 or 10000.0. Default value is 1000.0.

• discretize.sparse.max_buckets: Controls the maximum number of discretized values. Typical range is [10, 250]. Default value is 200. Meant for being used with sparse data.

  If you want to try a larger value up to 65000, then you need to edit include/header.h and replace using disc_sparse_value_t=unsigned char; by using disc_sparse_value_t=unsigned short;. However, this will increase the memory usage.

• discretize.dense.max_buckets: Controls the maximum number of discretized values. Typical range is [10, 65000]. Default value is 65000. Meant for being used with dense data.

• discretize.sparse.min_bucket_weights: Controls the minimum number of effective samples for each discretized value. Default value is 5.0. Meant for being used with sparse data.

• discretize.dense.min_bucket_weights: Controls the minimum number of effective samples for each discretized value. Default value is 5.0. Meant for being used with dense data.

• discretize.sparse.lamL2: Controls the degree of L2 regularization for discretization. Default value is 2.0. Meant for being used with sparse data.

• discretize.dense.lamL2: Controls the degree of L2 regularization for discretization. Default value is 2.0. Meant for being used with dense data.

• discretize.sparse.max_features: Controls the maximum number of selected features. Typical range is [1000, 10000000]. Default value is 80000. Meant for being used with sparse data.

• discretize.sparse.min_occurrences: Controls the minimum number of occurrences for a feature to be selected. Default value is 5. Meant for being used with sparse data.

• FastRGF crashes if training dataset is too small (#data < 28). rgf#92
• FastRGF does not provide any built-in method to calculate feature importances. rgf#109

Please post an issue at GitHub repository for any errors you encounter.

FastRGF is distributed under the MIT license. Please read the file LICENSE.

[1] Rie Johnson and Tong Zhang. Learning Nonlinear Functions Using Regularized Greedy Forest. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(5):942-954, May 2014.

[2] Ping Li. Robust LogitBoost and Adaptive Base Class (ABC) LogitBoost. UAI, 2010.

[3] Zhaohui Zheng, Hongyuan Zha, Tong Zhang, Olivier Chapelle, Keke Chen, Gordon Sun. A General Boosting Method and its Application to Learning Ranking Functions for Web Search. NIPS, 2007.