using the trained model. Each of the classifiers attempted to group the participants in the pedestrian
experiment by their personality characteristics as a
function of decision times to cross, to understand
the impact of the different displays, which was the
primary design question for this experiment.
Figure 4 illustrates representative trees from the three
different classification methods. Each method classifies
a pedestrian into two classes, such that Class 0 represents a cluster of pedestrians for whom the display on
the car (figure 2) did not matter and Class 1 represents
a cluster of pedestrians who leveraged the information
from the car’s display while making a crossing decision.
The designations of E, O, C, and A in figure 4 stand
for extraversion, openness, conscientiousness, and
agreeableness, which were the dominant personality
traits of those participating in the experiment.
The three decision tree models vary in terms of
tree structure and the type of variables used to clus-
ter the data. For instance, the decision tree formed
using FFT shows that just simply being above the
51-percent (median) threshold on the extraversion
personality scale predicted the use of one of the
displays in figure 2, a finding not revealed by the
hypothesis-driven ANOVA. The CART approach fur-
ther subdivided those people between agreeableness
and openness into Class 1 (people who depended on
the display). In comparison, the evolutionary tree
tended to group participants similarly to the FFT and
classified participants that relied on the display as
primarily extraverted (a score ≥ 49 or ≤ 53), but this
is difficult for some to understand, as it appears that
the openness root node is the primary relationship.
This process of training the model was repeated
1000 times for each classifier. Prediction accuracy was
calculated by averaging the number of correct predic-
tions at each iteration for every classifier, and was 62
percent for FFT, 67 percent for CART, and 51 percent
for evolutionary trees. Only the first three nodes of
the trained decision tree models were used to find
Classification Algorithm Advantages Disadvantages
FFT 1. Computationally fast, compared with
all the above-mentioned decision tree
1. Does not use all possible cues and
does not integrate information
while building decision trees.
Efficient and simple heuristic for
classification tasks, inspired by
2. Resultant decision trees are robust
and less susceptible to overfitting.
2. Because the heuristic computes no
utility or probability to quantify
the goodness of a branch split, it
may lead to nonoptimal splits.
CART 1. No underlying assumptions about
the nature of the observations
(for example, to be independent and
1. Possibility of nonoptimal splits
when learning a problem with
strong interdependency among the
Most commonly used classification
2. Results are invariant to the monotone
transformation of the predictor
variables (for example, squaring a
variable won’t change the structure
of the decision tree).
2. Unstable decision trees; small variations in the training data set can
lead to different tree structures.
3. Resultant decision trees are not
sensitive to outliers.
Evolutionary Trees 1. Best suited for problems where
multiple (locally optimal) solutions
are needed; cases where the best
solution may not always be realizable.
1. Computationally expensive and
large memory requirements.
A globally optimal classification
tree built using an evolutionary
2. Useful for problems with a huge
search space (for example, finding
optimal decision trees that are
2. Random nature of the algorithm
can yield different tree structures
with the same evaluation function
3. Large number of parameters
(crossover probability, mutation
rate, number of generations, and
so forth) that need to be manually
tuned, mostly by a trial-and-error
Table 1. Summary of the Advantages and Disadvantages of Popular Decision Tree Classification Algorithms.
NP, nondeterministic polynomial-time.