Generate, Visualise, and Evaluate Fast-and-Frugal Decision Trees

Create, visualize, and test fast-and-frugal decision trees (FFTs). FFTs are very simple decision trees for binary classification problems. FFTs can be preferable to more complex algorithms because they are easy to communicate, require very little information, and are robust against overfitting.

• FFTrees is an R package to create and visualize fast-and-frugal decision trees (FFTs) like the one below that predicts heart disease.

• Additional information about FFTs, and the FFTrees package can be found at Phillips, Neth, Woike & Gaissmaier, 2017. For seminal papers on FFTs, consult Martignon, Katsikopoulos & Woike, 2008 and Martignon, Vitouch, Takezawa & Forster, 2003

• A web-based Shiny version of FFTrees can be found at ShinyFFTrees

# Create an FFTrees object from the heartdisease data
heart.fft <- FFTrees(formula = diagnosis ~., 
                       data = heartdisease)
                       
# Plot the best tree
plot(heart.fft,
     main = "Heart Disease", 
     decision.labels = c("Healthy", "Disease"))

Package updates

1.3.5

• Bug fixes.

1.3.4

• Added class probability predictions with predict.FFTrees(type = "prob")

• Updated print.FFTrees() to display FFT #1 'in words' (from the inwords(x) function)

1.3.3

• Added show.X arguments to plot.FFTrees() that allow you to selectively turn on or turn off elements when plotting an FFTrees object.

• Added label.tree, label.performance arguments to plot.FFTrees() that allow you to specify plot (sub) labels.

• Bug fixes

  • Issues when passing an existing FFTrees object to a new call to FFTrees().

1.3.0

• Many additional vignettes (e.g.; Accuracy Statistics and Heart Disease Tutorial) and updates to existing vignettes.

• Added cost.outcomes and cost.cues to allow the user to specify specify the cost of outcomes and cues. Also added a new cost statistic throughout outputs.

• Added inwords(), a function that converts an FFTrees object to words.

• Added my.tree argument to FFTrees() that allows the user to specify an FFT verbally. E.g., my.tree = 'If age > 30, predict True. If sex = {m}, predict False. Otherwise, predict True'.

• Added positive predictive value ppv, negative predictive value npv and balanced predictive value bpv as primary accuracy statistics throughout.

• Added support for two FFT construction algorithms from Martignon et al. (2008): "zigzag" and "max". The algorithms are contained in the file heuristic_algorithm.R and can be implemented in FFTrees() as arguments to algorithm.

1.2.3

• Added sens.w argument to allow differential weighting of sensitivities and specificities when selecting and applying trees.

• Fixed bug in calculating importance weightings from FFForest() outputs.

1.2.0

• Changed wording of statistics throughout package. hr (hit rate) and far (false alarm rate) are now sens for sensitivity, and spec for specificity (1 - false alarm rate)

• The rank.method argument is now depricated. Use algorithm instead.

• Added stats argument to plot.FFTrees(). When stats = FALSE, only the tree will be plotted without reference to any statistical output.

• Grouped all competitive algorithm results (regression, cart, random forests, support vector machines) to the new x.fft$comp slot rather than a separate first level list for each algorithm. Also replaced separate algorithm wrappers with one general comp.pred() wrapper function.

• Added FFForest(), a function for creating forests of ffts, and plot.FFForest(), for visualizing forests of ffts. This function is very much still in development.

• Added random forests and support vector machines for comparison in FFTrees() using the randomForest and e1071 packages.

• Changed logistic regression algorithm from the default glm() version to glmnet() for a regularized version.

• predict.FFTrees() now returns a vector of predictions for a specific tree rather than creating an entirely new FFTrees object.

• You can now plot cue accuracies within the plot.FFTrees() function by including the plot.FFTrees(what = 'cues') argument. This replaces the former showcues() function.

• Many cosmetic changes to plot.FFTrees() (e.g.; gray levels, more distinct classification balls). You can also control whether the results from competing algorithms are displayed or not with the comp argument.

• Bug-fixes

  • Fixed a bug where levels with no classifications are not plotted correctly.

1.1.7

• Trees can now use the same cue multiple times within a tree. To do this, set rank.method = "c" and repeat.cues = TRUE.

• Bug-fixes

  • You can (and should!) now have a column of NAs for the criterion in test datasets to represent data where the criterion is unknown.
  • FFTrees() now supports a single predictor (e.g.; formula = diagnosis ~ age) which previously did not work.

1.1.6

• Streamlined code to improve cohesion between functions. This may cause issues with FFTrees objects created with earlier versions of the package. They will need to be re-created.

• Updated, clearer print.FFTrees() method to see important info about an FFTrees object in matrix format.

• Training and testing statistics are now always in seperate objects (e.g.; data$train, data$test) to avoid confusion.

• Bug-fixes

  • predict.FFTrees() now works much better by passing a new dataset (data.test) as a test dataset for an existing FFTrees object.

1.1.5

• Bug-fixes
  • Plotting parameters mar and layout are now reset after running plot.FFTrees()

1.1.4

• Bug-fixes

  • Plotting no longer fails when there is only one branch in the tree.
  • Changed which.tree argument in plot.FFTrees() to tree to conform to blog posts.
  • predict.FFTrees() now works better with tibble inputs.

• Changed the fft label to FFTrees throughout the package to avoid confusion with fast fourier transform. Thus, the main tree building function is now FFTrees() and the new tree object class is FFTrees