The package ccf implements canonical correlations forests (CCFs) for use inside R. These present a novel classification algorithm for machine learning tasks, that are often able outperform common methods for predictive classifiers. The CCF algorithm is based on an ensemble of decision trees together with a canonical correlation analysis. The purpose is to de-correlated individual trees and thus improve the predictive performance.

Brief summary of CCF algorithm

A decision tree is a predictive model that sequentially divides the input space, each for which a local classification and regression model is calculated (e.g. with a simple majority vote). Thereby, it generates a tree-like structure, whose leaves usually group data points belonging ideally to the same class. One can often achieve a better performance by combining individual trees and average over them. This is known as a decision forest or random forest.

A canonical correlation forest is a now tree ensemble method. While the concept is similar to a forest, its specific characteristics often achieve a favorable predictive performance. It trains the trees by using a canonical correlation analysis (CCA) in order to find a feature projection that gives a maximal correlation between features. It then chooses the best split in this projected space.

For a thorough explanation and derivation refer to:

• Rainforth, T., and Wood, F. (2015): Canonical correlation forest, arXiv preprint, arXiv:1507.05444.

Overview

The most important functions in ccf are:

• canonical_correlation_forest() compute classifier based on canonical correlation forests. It supports both a matrix-like input, as well as the common convention using a formula.

• predict() applies classifier to unseen data and predicts the class outcome.

• plot visualizates the underlying decision surface.

To see examples of these functions in use, check out the help pages, the demos and this README (which is identical to the vignette).

Installation

Using the devtools package, you can easily install the latest development version of ccf with

install.packages("devtools")

# Option 1: download and install latest version from ‘GitHub’
devtools::install_github("jandob/ccf")

# Option 2: install directly from bundled archive
# devtoos::install_local("ccf_0.1.0.tar.gz")

Notes:

• In the case of option 2, you have to specify the path either to the directory of ccf or to the bundled archive ccf_1.0.0.tar.gz

• A CRAN version has not yet been released, but we are working on it. This also applies to the integration into predictive frameworks such as caret or mlr.

Usage

This section shows the basic functionality of how to train a canonical correlation forests and make predictions based on it. First, load the corresponding package ccf.

library(ccf)

The interface follows common R conventions as used by other machine learning routines. Therefore, the usage is fairly straightforward.

# load sample dataset
data(spirals)

d_train <- spirals[1:1000, ]
d_test <- tail(spirals, 1000)

# compute classifier on training data
## variant 1: matrix input
m1 <- canonical_correlation_forest(d_train[, c("x", "y")], d_train$class, ntree = 10)
## variant 2: formula notation
#m2 <- canonical_correlation_forest(class ~ ., d_train)

# compute predictive accuracy
#get_missclassification_rate(m1, d_test)
#get_missclassification_rate(m2, d_test)

# plot the decision surface of the classifier
ccf_plot <- plot_decision_surface(
  m1, d_test[, c("x", "y")], d_test$class, title = "CCF with 20 trees")

License

ccf is released under the MIT License

Copyright (c) 2016 Janosch Dobler & Stefan Feuerriegel