FFTrees 1.7.5

The R package FFTrees creates, visualizes and evaluates fast-and-frugal decision trees (FFTs) for solving binary classification tasks following the methods described in Phillips, Neth, Woike & Gaissmaier (2017, as html | PDF).

What are fast-and-frugal trees (FFTs)?

Fast-and-frugal trees (FFTs) are simple and transparent decision algorithms for solving binary classification problems. The key feature making FFTs faster and more frugal than other decision trees is that every node allows for a decision. When predicting new outcomes, the performance of FFTs competes with more complex algorithms and machine learning techniques, such as logistic regression (LR), support-vector machines (SVM), and random forests (RF). Apart from being faster and requiring less information, FFTs tend to be robust against overfitting, and easy to interpret, use, and communicate.

Installation

To install the latest release of FFTrees from CRAN evaluate:

install.packages("FFTrees")

The current development version of FFTrees can be installed from GitHub with:

# install.packages("devtools")
devtools::install_github("ndphillips/FFTrees", build_vignettes = TRUE)

Getting started

As an example, let’s create a FFT predicting heart disease status (Healthy vs. Diseased) based on the heartdisease dataset included in FFTrees:

library(FFTrees)  # load package

Using data

The heartdisease data provides medical information for 303 patients that were tested for heart disease. The full data were split into two subsets: A heart.train dataset for fitting decision trees, and heart.test dataset for a testing the resulting trees. Here are the first rows and columns of both subsets of the heartdisease data:

• heart.train (the training / fitting dataset) contains the data from 150 patients:

head(heart.train)
#> # A tibble: 6 × 14
#>   diagnosis   age   sex cp    trestbps  chol   fbs restecg thalach exang oldpeak
#>   <lgl>     <dbl> <dbl> <chr>    <dbl> <dbl> <dbl> <chr>     <dbl> <dbl>   <dbl>
#> 1 FALSE        44     0 np         108   141     0 normal      175     0     0.6
#> 2 FALSE        51     0 np         140   308     0 hypert…     142     0     1.5
#> 3 FALSE        52     1 np         138   223     0 normal      169     0     0  
#> 4 TRUE         48     1 aa         110   229     0 normal      168     0     1  
#> 5 FALSE        59     1 aa         140   221     0 normal      164     1     0  
#> 6 FALSE        58     1 np         105   240     0 hypert…     154     1     0.6
#> # … with 3 more variables: slope <chr>, ca <dbl>, thal <chr>

• heart.test (the testing / prediction dataset) contains data from a new set of 153 patients:

head(heart.test)
#> # A tibble: 6 × 14
#>   diagnosis   age   sex cp    trestbps  chol   fbs restecg thalach exang oldpeak
#>   <lgl>     <dbl> <dbl> <chr>    <dbl> <dbl> <dbl> <chr>     <dbl> <dbl>   <dbl>
#> 1 FALSE        51     0 np         120   295     0 hypert…     157     0     0.6
#> 2 TRUE         45     1 ta         110   264     0 normal      132     0     1.2
#> 3 TRUE         53     1 a          123   282     0 normal       95     1     2  
#> 4 TRUE         45     1 a          142   309     0 hypert…     147     1     0  
#> 5 FALSE        66     1 a          120   302     0 hypert…     151     0     0.4
#> 6 TRUE         48     1 a          130   256     1 hypert…     150     1     0  
#> # … with 3 more variables: slope <chr>, ca <dbl>, thal <chr>

Most of the variables in our data are potential predictors. The criterion variable is diagnosis — a logical column indicating the true state for each patient (TRUE or FALSE, i.e., whether or not the patient suffers from heart disease).

Creating fast-and-frugal trees (FFTs)

Now let’s use FFTrees() to create FFTs for the heart.train data and evaluate their predictive performance on the heart.test data:

• Create an FFTrees object from the heartdisease data:

# Create an FFTrees object from the heartdisease data: 
heart.fft <- FFTrees(formula = diagnosis ~., 
                     data = heart.train,
                     data.test = heart.test, 
                     decision.labels = c("Healthy", "Disease"))
#> Setting 'goal = bacc'
#> Setting 'goal.chase = bacc'
#> Setting 'goal.threshold = bacc'
#> Setting cost.outcomes = list(hi = 0, mi = 1, fa = 1, cr = 0)
#> Growing FFTs with ifan:
#> Fitting other algorithms for comparison (disable with do.comp = FALSE) ...

• Printing an FFTrees object shows basic information and summary statistics (on the best training tree, FFT #1):

# Print:
heart.fft
#> FFTrees 
#> - Trees: 7 fast-and-frugal trees predicting diagnosis
#> - Outcome costs: [hi = 0, mi = 1, fa = 1, cr = 0]
#> 
#> FFT #1: Definition
#> [1] If thal = {rd,fd}, decide Disease.
#> [2] If cp != {a}, decide Healthy.
#> [3] If ca > 0, decide Disease, otherwise, decide Healthy.
#> 
#> FFT #1: Training Accuracy
#> Training data: N = 150, Pos (+) = 66 (44%) 
#> 
#> |          | True + | True - | Totals:
#> |----------|--------|--------|
#> | Decide + | hi  54 | fa  18 |      72
#> | Decide - | mi  12 | cr  66 |      78
#> |----------|--------|--------|
#>   Totals:        66       84   N = 150
#> 
#> acc  = 80.0%   ppv  = 75.0%   npv  = 84.6%
#> bacc = 80.2%   sens = 81.8%   spec = 78.6%
#> 
#> FFT #1: Training Speed, Frugality, and Cost
#> mcu = 1.74,  pci = 0.87,  E(cost) = 0.200

• To evaluate the predictive performance of an FFT, we plot an FFTrees object to visualize a tree and its performance (on the test data):

# Plot the best tree applied to the test data: 
plot(heart.fft,
     data = "test",
     main = "Heart Disease")

Figure 1: A fast-and-frugal tree (FFT) predicting heart disease for test data and its performance characteristics.

• Additionally, we can compare the predictive performance between different machine learning algorithms on a range of metrics:

# Compare predictive performance across algorithms: 
heart.fft$competition$test
#> # A tibble: 5 × 17
#>   algorithm     n    hi    fa    mi    cr  sens  spec    far   ppv   npv   acc
#>   <chr>     <int> <int> <int> <int> <int> <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>
#> 1 fftrees     153    64    19     9    61 0.877 0.762 0.238  0.771 0.871 0.817
#> 2 lr          153    55    13    18    67 0.753 0.838 0.162  0.809 0.788 0.797
#> 3 cart        153    50    19    23    61 0.685 0.762 0.238  0.725 0.726 0.725
#> 4 rf          153    59     8    14    72 0.808 0.9   0.1    0.881 0.837 0.856
#> 5 svm         153    55     7    18    73 0.753 0.912 0.0875 0.887 0.802 0.837
#> # … with 5 more variables: bacc <dbl>, wacc <dbl>, cost <dbl>,
#> #   cost_decisions <dbl>, cost_cues <dbl>

Building FFTs from verbal descriptions

Because fast-and-frugal trees are so simple, we even can create FFTs ‘from words’ and apply them to data!

For example, let’s create a tree with the following three nodes and evaluate its performance on the heart.test data:

1. If sex = 1, predict Disease.
2. If age < 45, predict Healthy.
3. If thal = {fd, normal}, predict Healthy,
   otherwise, predict Disease.

These conditions can directly be supplied to the my.tree argument of FFTrees():

# Create custom FFT 'in words' and apply it to test data:

# 1. Create my own FFT (from verbal description):
my.fft <- FFTrees(formula = diagnosis ~., 
                  data = heart.train,
                  data.test = heart.test, 
                  decision.labels = c("Healthy", "Disease"),
                  my.tree = "If sex = 1, predict Disease.
                             If age < 45, predict Healthy.
                             If thal = {fd, normal}, predict Healthy,  
                             Otherwise, predict Disease.")

# 2. Plot and evaluate my custom FFT (for test data):
plot(my.fft,
     data = "test",
     main = "My custom FFT")

Figure 2: An FFT predicting heart disease created from a verbal description.

As we can see, this particular tree is somewhat biased: It has nearly perfect sensitivity (i.e., is good at identifying cases of Disease) but suffers from low specificity (i.e., is not so good at identifying Healthy cases). Overall, the accuracy of our custom tree exceeds the data’s baseline by a fair amount. However, exploring FFTrees further will quickly enable you to design much better FFTs.

References

We had a lot of fun creating FFTrees and hope you like it too! As a comprehensive, yet accessible introduction to FFTs, we recommend reading our article in the journal Judgment and Decision Making (2017, volume 12, issue 4), entitled FFTrees: A toolbox to create, visualize,and evaluate fast-and-frugal decision trees (available in html | PDF ).

Citation (in APA format):

• Phillips, N. D., Neth, H., Woike, J. K. & Gaissmaier, W. (2017). FFTrees: A toolbox to create, visualize, and evaluate fast-and-frugal decision trees. Judgment and Decision Making, 12 (4), 344–368. Retrieved from https://journal.sjdm.org/17/17217/jdm17217.pdf

We encourage you to read the article to learn more about the history of FFTs and how the FFTrees package creates, visualizes, and evaluates them. When using FFTrees in your own work, please cite us and share your experiences (e.g., on GitHub) so we can continue developing the package.

Here are some scientific publications that have used FFTrees (see Google Scholar for the full list):

• Lötsch, J., Haehner, A., & Hummel, T. (2020). Machine-learning-derived rules set excludes risk of Parkinson’s disease in patients with olfactory or gustatory symptoms with high accuracy. Journal of Neurology, 267(2), 469-478.

• Kagan, R., Parlee, L., Beckett, B., Hayden, J. B., Gundle, K. R., & Doung, Y. C. (2020). Radiographic parameter-driven decision tree reliably predicts aseptic mechanical failure of compressive osseointegration fixation. Acta Orthopaedica, 91(2), 171-176.

• Klement, R. J., Sonke, J. J., Allgäuer, M., Andratschke, N., Appold, S., Belderbos, J., … & Mantel, F. (2020). Correlating dose variables with local tumor control in stereotactic body radiotherapy for early stage non-small cell lung cancer: A modeling study on 1500 individual treatments. International Journal of Radiation Oncology * Biology * Physics.

• Nobre, G. G., Hunink, J. E., Baruth, B., Aerts, J. C., & Ward, P. J. (2019). Translating large-scale climate variability into crop production forecast in Europe. Scientific Reports, 9(1), 1-13.

• Buchinsky, F. J., Valentino, W. L., Ruszkay, N., Powell, E., Derkay, C. S., Seedat, R. Y., … & Mortelliti, A. J. (2019). Age at diagnosis, but not HPV type, is strongly associated with clinical course in recurrent respiratory papillomatosis. PloS One, 14(6).

[File README.Rmd last updated on 2022-09-15.]