Classification and Regression Trees

This software uses regression trees inside a Random Forest to classify matrices of data. There are two versions of the software: a Python version in the python folder. There is a C++ version which is faster and more accurate in the root folder. Both versions employ parallel programming and run in multiple threads or processes.

Running the program

To run the program, use the following parameters: ./randomForest -t training_file -c data_to_classify -p 16 -n 1000 -f 30 -m 0.1

Required Parameters

• -t training_file
  • input matrix file to train trees
• -s filename to write soft labels
  • output of softlabels for output to AUPR program

Optional Parameters

• -p 16
  • use 16 threads in a thread pool to process trees
• -n 1000
  • use 1000 trees in the forest
• -f 30
  • use a subset of 30 features for each tree
• -g 0.1
  • minimum gain of sum of squares to continue splitting
  • gain = sum_of_squares - (left_squares + right_squares)
• -m

Class Descriptions

• Matrix - 2d double data structure with utility functions
• Classifier - Base class for all classifiers
  • TreeNode - Individual node for a regression tree including the root node.
  • Forest - Single-threaded Regression Forest Classifier
    • ParallelForest - Multi-threaded training using thread pools

Other Code Files

• main.cpp - main function and argument processing
• stats.cpp/hpp - statistical functions (mean, mode, variance, regression, etc.)
• util.cpp/hpp - utility functions similar to Python's built-ins
• pthread_pool.c/h - thread pool C code from the Internet

TODO

Migrate away from c thread_pool code.  Redo the code in an object-oriented manner.