Micro-Pattern Detection and Analysis in Gaze Data via Mathematical Optimization and Machine Learning

Liu, Wen

Etd

Micro-Pattern Detection and Analysis in Gaze Data via Mathematical Optimization and Machine Learning

Public

The use of eye tracking analysis to understand human behavior and cognition is increasingly prevalent in user experience research. Eye gaze data consists of a sequence of eye movement events, such as fixation and saccade, which can be used to analyze focus of attention and awareness under a variety of visual stimuli. The distribution of gaze points within individual fixations, which we call micro-patterns, has to date been largely unexplored. This work uses mathematical optimization and machine learning to explore micro-patterns in gaze data, and thereby improve the fundamental unit of analysis for attention and awareness in eye-tracking studies.\n\nThe result is enhanced accuracy of location and level of attention intensity. The primary research is to study micro-patterns in gaze data by developing fixation detection algorithms using data science technologies. Fixation inner-density (FID), introduced for the first time in this dissertation, measures the compactness of a fixation. It exhibits significant information about focused attention and effort. In Chapter 1, integer optimization and algorithmic techniques are combined to identify fixations in gaze point sequences by optimizing for inner-density. The computational results in Chapter 1 together with the experiments in Chapter 2, demonstrate that this approach, also known as the FID filter, outperforms methods used in existing commercial eye trackers in fixation refinement. Moreover, it has great potential to contribute to user experience research by providing better representation of attention and awareness, which is the fundamental unit of analysis in behavioral studies. We further extend this research in two dimensions. The first extension, known as the FID+ filter, advances the integer optimization techniques to identify fixation outliers in gaze point sequences. As introduced in Chapter 3, this enhances the FID filter by accounting for outlier sensitivity. The second extension is a set of experiments to explore the automated recommendation of the density intensity modulation parameter α to the FID filter users. Chapter 4 discusses current findings from the experiments of recommending suitable α levels on how two eye-tracking datasets were manually labeled, and experimental findings on recommending suitable α levels.\n\nThese developments serve as fundamental building blocks for a real-time system for gaze fixation detection using inner-density. Such a system can provide instant and accurate gaze analysis, and thereby enable the ability to provide immediate feedback to the user. This may have significant implications and expand the application scope of eye tracking, and will be beneficial to Human Computer Interaction and behavioral research through the development of innovative and personalized user experiences.

Creator