Example: Heart Disease
First, we’ll calculate a classification threshold for each cue using cuerank():
heartdisease.ca <- FFTrees::cuerank(formula = diagnosis ~.,
data = heartdisease)
# Print key results
heartdisease.ca[c("cue", "threshold", "direction", "bacc")]## cue threshold direction bacc
## 1 age 55 > 0.6347166
## 2 sex 0 > 0.6295841
## 3 cp a = 0.7587954
## 4 trestbps 140 > 0.5579707
## 5 chol 242 > 0.5677092
## 6 fbs 0 > 0.5090147
## 7 restecg hypertrophy,abnormal = 0.5881953
## 8 thalach 148 < 0.7042902
## 9 exang 0 > 0.7032593
## 10 oldpeak 0.8 > 0.6978856
## 11 slope flat,down = 0.6936743
## 12 ca 0 > 0.7308738
## 13 thal rd,fd = 0.7596508Here, we see the best decision threshold for each cue that maximizes its balanced accuracy (bacc) when applied to the entire dataset (independently of other cues). For example, for the age cue, the best threshold is age > 55 which leads to a balanced accuracy of 0.63. In other words, if we only had the age cue, then the best decision is: “If age > 55, predict heart disease, otherwise, predict no heart disease”.
Let’s confirm that this threshold makes sense. To do this, we can plot the bacc value for all possible thresholds as in Figure @ref(fig:agethreshold):
Plotting the balanced accuracy of each decision threshold for the age cue.
Next, the cues are ranked by their balanced accuracy. Let’s do that with the heart disease cues:
# Rank heartdisease cues by balanced accuracy
heartdisease.ca <- heartdisease.ca[order(heartdisease.ca$bacc, decreasing = TRUE),]
# Print the key columns
heartdisease.ca[c("cue", "threshold", "direction", "bacc")]## cue threshold direction bacc
## 13 thal rd,fd = 0.7596508
## 3 cp a = 0.7587954
## 12 ca 0 > 0.7308738
## 8 thalach 148 < 0.7042902
## 9 exang 0 > 0.7032593
## 10 oldpeak 0.8 > 0.6978856
## 11 slope flat,down = 0.6936743
## 1 age 55 > 0.6347166
## 2 sex 0 > 0.6295841
## 7 restecg hypertrophy,abnormal = 0.5881953
## 5 chol 242 > 0.5677092
## 4 trestbps 140 > 0.5579707
## 6 fbs 0 > 0.5090147Now, we can see that the top five cues are thal, cp, ca, thalach and exang. Because ffts rarely exceed 5 cues, we can expect that the trees will use a subset (not necessarily all) of these 5 cues.
We can also plot the cue accuracies in ROC space using the showcues() function:
# Show the accuracy of cues in ROC space
showcues(cue.accuracies = heartdisease.ca)
Cue accuracies for the heartdisease dataset. The top 5 cues in terms of balanced accuracy are highlighted.
Next, grow.FFTrees() will grow several trees from these cues using different exit structures:
# Grow FFTs
heartdisease.ffts <- grow.FFTrees(formula = diagnosis ~.,
data = heartdisease)
# Print the tree definitions
heartdisease.ffts$tree.definitions## tree cues nodes classes exits thresholds directions
## 1 1 thal;cp;ca;thalach 4 c;c;n;n 0;0;0;0.5 rd,fd;a;0;148 =;=;>;<
## 5 2 thal;cp;ca;thalach 4 c;c;n;n 0;0;1;0.5 rd,fd;a;0;148 =;=;>;<
## 3 3 thal;cp;ca 3 c;c;n 0;1;0.5 rd,fd;a;0 =;=;>
## 2 4 thal;cp;ca 3 c;c;n 1;0;0.5 rd,fd;a;0 =;=;>
## 6 5 thal;cp;ca;thalach 4 c;c;n;n 1;0;1;0.5 rd,fd;a;0;148 =;=;>;<
## 4 6 thal;cp;ca;thalach 4 c;c;n;n 1;1;0;0.5 rd,fd;a;0;148 =;=;>;<
## 7 7 thal;cp;ca;thalach 4 c;c;n;n 1;1;1;0.5 rd,fd;a;0;148 =;=;>;<Here, we see that we have 7 different trees, each using some combination of the top 5 cues we identified earlier. For example, tree 1 uses the top 4 cues, while tree 3 uses only the top 3 cues. Why is that? The reason is that the algorithm also prunes lower branches of the tree if there are too few cases classified at lower levels. By default, the algorithm will remove any lower leves that classfify fewer than 10% of the original cases. The pruning criteria can be controlled using the stopping.rule, stopping.par and max.levels arguments in grow.FFTrees()
Now let’s use the wrapper function FFTrees() to create the trees all at once. We will then plot tree #4 which, according to our results above, should contain the cues thal, cp, ca"
library(FFTrees)
# Create trees
heart.fft <- FFTrees(formula = diagnosis ~., data = heartdisease)
# Plot tree # 4
plot(heart.fft,
stats = FALSE, # Don't include statistics
tree = 4)
Alternative algorithms
algorithm = "c": The"c"algorithm is identical to the"m"algorithm with one (important) exception: Inalgorithm = "c", the thresholds and rankings of cues are recalculated for each level in the FFT conditioned on the exemplars that were not classified at higher leves in the tree. For example, in theheartdiseasedata, usingalgorithm = "c"would first classify some cases using thethalcue at the first level, and would then calculate new accuracies for the remaining cues on the remaining cases that were not yet classified. This algorithm is appropriate for datasets where cue validities systematically differ for different (and predictable) subsets of data. However, because it calculates cue thresholds for increasingly smaller samples of data as the tree grows, it is also, potentially, more prone to overfitting compared toalgorithm = "m"