This work is public research, published in peer-reviewed journals and recognised with a Best Paper Award. It is the longest-running project in our portfolio, and the clearest example of research that did not stop at publication.
How it works
Record where people look
Eye tracking on expert and trainee, same moving radar picture.
Treat a scan as a path
The targets move, so a scan becomes transitions through a network.
Measure distance from expert
Graph clustering scores how far a trainee diverges.
Put it in front of instructors
Two systems bring it into the training room, live.
Comparing how two people look at a moving world
Expert air traffic controllers scan a radar display differently from trainees. Everyone in the training community knew this. Nobody could express it in a way an instructor could teach against, because the expertise is tacit — experienced controllers cannot fully articulate what they are doing, and watching them does not reveal it either.
Eye tracking produces the raw material, but not the answer. The obvious approach — comparing where two people looked — fails immediately, because the objects being looked at are themselves moving. A gaze point means nothing in isolation; what matters is which aircraft was being attended to, in what order, relative to everything else on the screen at that instant. Standard similarity measures have nothing to say about sequences over dynamic, multi-element scenes.
The problem was never the eye tracking. It was that no one had a way to say two scanning patterns were similar.
So we built one. We designed a dynamic network representation of eye-movement data over moving objects, and a graph-clustering algorithm using a modified n-gram distance combined with graph modularity — treating a scan pattern as a sequence of transitions through a network rather than a cloud of points. Domain expertise was incorporated directly into the models to keep pattern recognition consistent, which is what made the resulting comparisons stable enough to benchmark against.
Two systems, in the training rooms
A benchmark that lives in a paper changes nothing. The second half of this work was engineering and deploying two custom software systems for the FAA that brought the methodology into operational controller training — so an instructor could see how a trainee's attention pattern diverged from the expert benchmark, during training, rather than inferring it from outcomes afterwards.
This is the part most research does not reach. It required the method to survive contact with real training schedules, real instructors who were not researchers, and the ordinary operational constraints of an aviation training environment.
If you're facing something similar
Where expertise exists but cannot be articulated, asking the expert is the wrong instrument. Measure the behaviour instead, and the thing people could not describe becomes something you can benchmark, teach against and improve — whether the expert is a controller, a claims adjuster, a plant operator or a surgeon.
The second lesson is about the gap between a method and a result. The award was for the algorithm; the 60% came from the two systems that put it in front of instructors. Those are different pieces of work, and only the second one changed anything.