Decision Tree Classifier
B The Traditional Hypothesis-driven
Approach The Bottom-Up ML Approach
differences, which is an important contextual cue for
the people in our analysis, as opposed to aggregating the data across an expected population. In typical hypothesis-driven statistical analyses, individual
differences are treated as uncontrollable variability.
The use of blocked designs, covariance analyses, and
other related pre- and post hoc tools attempt to partition and minimize the effect of individual differences, but doing so potentially causes researchers to
lose important and useful information. We hypothesize that by using ML to preprocess the data, we can
actually identify and leverage individual differences
to account for one source of context.
Using the multistage approach depicted in figure
5b to reanalyze the pedestrian experiment results
described previously with CART, the new results
showed that some pedestrians do leverage the information from their surroundings including the external display, as opposed to just relying on legacy
behaviors suggested by the traditional approach
(figure 5a). Of the original 55 subjects, 42 percent
of participants using CART predominantly relied on
the displays more than any other factor such as speed
of the car, or which side of the road they were on.
Given the novelty of self-driving cars and the
fact that people have not had much exposure to
new forms of vehicle-to-pedestrian communication,
It is crucial when designing autonomous technol-
ogies to consider carefully how such systems can
effectively interact with both operators and other
relevant stakeholders, particularly in safety-critical
(a) Treating independent variables as factors and performing inferential omnibus tests. (b) Clustering people based on their individual
attributes like personality, age, and crossing position (cross-walker versus jay-walker), in terms of which vehicle attributes they focused on
more (display, speed, or direction of the approaching car), and then performing statistical tests on these groups.