Research | Home

Research experience

I've applied my background in data science and machine learning to address human factor challenges. The following sections describe some of the research projects I've been involved in.

I've participated in numerous projects whose goal was to investigate the visual scanning strategies and decision-making of expert air traffic controllers through a combination of eye tracking and 1:1 interviews.

A visual scanning pattern of an expert en-route controller searching for a potential conflict.

Check out this paper

A heatmap showcasing the accumulation of expert tower controllers eye fixations' over after managing incoming and departing aircraft in a simulator (red indicates more eye fixations).

Check out this paper

I've also explored using biometric data, specifically eye movements and brain oxygenation, to investigate the cognitive workload of expert tower controllers.

In the study, we saw how the oxygenation at the pre-frontal cortex, as well as eye movement metrics, such as the number of eye fixations, increased as the task load -- the number of aircraft -- also increased.

Check out this paper

I've supported the development of multi-person virtual reality applications that use biometric data, namely eye movements and brain oxygenation, to discover smart-learning methodologies.

In one application, users can connect multiple objects together to create semantic networks, depicting how objects are linked together (e.g. by color)

As users interact with the application, their eye movement and brain oxygenation data is being collected (yellow dots indicate where users were observing at that point in time)

Check out this paper

I've compared the performance of cross-validation methods to train machine learning models using biometric data (e.g. eye movements) collected over time and from multiple users in order to predict cognitive workload.

A visualization depicting the cross-validation methods used (k-fold and population-informed forward chain) to train machine learning models (e.g. ridge regression) using the data collected from six participants.

The models trained using k-fold overestimated performance, as k-fold does not consider the time-series and user characteristics contained in the data.

Check out this paper

I've developed a framework to compare visual scan paths between multiple participants using complimentary similarity metrics. Namely the string-edit algorithm, which considers the order used to observe information, and the Jaccard similarity coefficient, which considers what information was observed.

An example of an expert tower controller's visual scan path prior to issuing a clearance. Using multiple similarity metrics can help us get a better picture of what makes this visual scan path similar (or more different) to other visual scan paths.

The framework uses the average string-edit and Jaccard coefficient similarity between visual scan paths to classify them.

Check out this paper