You can visualise fft objects in one of two ways: First, you can visualise cue accuracies with showcues()
. Second, you can visualise individual trees and peformance statistics by applying the plot()
function to an fft
object.
The titanic dataset contains survival statistics of passengers on the Titanic. For each passsenger, we know what passenger class (s)he was, his/her age, his/her sex, and whether or not (s)he survived. Our goal is to build a FFT that predicts whether or not a passenger will survive based on these cues.
First, let’s create an fft object called titanic.fft
from the titanic
dataset.
titanic.fft <- FFTrees(
formula = survived ~.,
data = titanic
)
You can visualise individual cue accuracies (specifically their hit rates and false alarm rates) using the showcues()
function. Let’s apply the function to the titanic.fft
object to see how accurate each of the cues were on their own in predicting survival:
showcues(titanic.fft,
main = "Titanic cue accuracy")
Wow. None of the cues did very well on their own. Good performing cues should be in the top left hand of the graph (i.e.; low false alarm rate and high hit rate). It looks like the best cue was sex
, followed by class
. age
was a pretty terrible cue.
To plot the tree from an fft object, use plot()
. You can add some stylistic arguments like description
and decision.names
. Let’s plot one of the trees:
plot(titanic.fft,
main = "Titanic",
decision.names = c("Died", "Survived"))
This plot contains a lot of information, here are the main elements:
The top row of the plot shows the main dataset information.
The middle row shows the tree as well as how many examples were classified at each level in the tree. For example, this tree could be understood as: “If the person is not male, predict they survived. Then, if the person is neither in first nor second class, predict they died. Finally, if the person is a child, predict they survived.”
The bottom row shows general performance statistics of the tree. If fitting data (i.e.; data used to build the tree) are displayed, you’ll see a “Fitting” label. If a testing dataset separate from the one used to build the tree is used, you’ll see a “Prediction” label. The classification table on the left side shows the relationship between tree decisions and the truth. CR (Correct Rejection) and H (Hit) are correct decisions. MI (Miss) and FA (False-alarm) are incorrect decisions. The next three levels show cumulative tree performance in terms of Specificity, Hit Rate, D-prime, and AUC (area under the curve). Finally, the plot on the right shows an ROC curve comparing the performance of all trees in the fft object. Additionally, the performance of logistic regression (blue) and CART (red) are shown. The tree plotted in the middle row is highlighted in a solid green color (i the case above, tree #3).
You can specify additional arguments to the plot()
command that will change what is displayed
tree
: Which tree do you want to plot? You can specify an integer such as tree = 2
will plot the tree #2 in the fft object, or tree = "best.train"
which will use the best training tree.
data
: Which data do you want to apply the tree to? You can specify data = "train"
or data = "test"
to use the training or testing datasets stored in the fft
object. Alternatively, you can specify a new dataset (e.g.; data = test.data
. If you specify a new dataset, the function will automatically apply the tree to the new data and calculate performance statitics (using the predict.FFTrees()
function).
For example, let’s repeat the previous analysis, but now we’ll create separate training and test datasets by including the train.p = .5
argument. This will split the dataset into a 50% training set, and a 50% testing set:
set.seed(100)
titanic.pred.fft <- FFTrees(formula = survived ~.,
data = titanic,
train.p = .5)
Here is the best training tree applied to the training data:
plot(titanic.pred.fft,
tree = "best.train",
main = "Titanic",
decision.names = c("Died", "Survived"))
The best training tree (tree #3) had a high specificity of 92%, but a low hit rate of just 48%. However, as we can see in the ROC table, LR didn’t perform much better, and CART did even worse than tree #3.
Now let’s apply the same tree to the test dataset:
plot(titanic.pred.fft,
tree = "best.train",
data = "test",
main = "Titanic",
decision.names = c("Died", "Survived"))
Performance has actually increased in this test data (e.g.; the hit-rate is up to 54%). However, both logistic regression and CART did similarly. Let’s see how tree #4, the most liberal tree, did:
plot(titanic.pred.fft,
tree = 4,
data = "test",
main = "Titanic",
decision.names = c("Died", "Survived"))
Tree #4 was able to increase the testing hit-rate up to 65%, but at a cost of a lower specificity of 70%.