Unsupervised parsing of gaze data with a beta-process vector auto-regressive hidden Markov model

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Data Science

Background:

• Eye-tracking data analysis typically involves classifying gaze data into fixations and saccades using predefined rules.
• Existing automated methods often rely on fixed thresholds or limited, predetermined eye movement categories.
• Recent advancements utilize time-varying, data-driven thresholds for more nuanced parsing.

Purpose of the Study:

• To present an alternative, data-driven approach for parsing eye-tracking data using a novel statistical model.
• To overcome limitations of existing methods by providing a statistical framework and flexible category identification.
• To demonstrate the efficacy of the proposed model on both high and low-sampling rate eye-tracking datasets.

Main Methods:

• Development and application of the beta-process vector auto-regressive hidden Markov model (BP-AR-HMM).
• Utilizing a latent process within the BP-AR-HMM to model the number and types of eye movements dynamically.
• Testing the BP-AR-HMM on diverse datasets, including high-sampling rate data (Andersson et al.) and low-sampling rate data (DIEM project).

Main Results:

• The BP-AR-HMM successfully identified over five distinct categories of eye movements.
• Identified movements included standard fixations and saccades, as well as potentially novel categories like post-saccadic oscillations and smooth pursuit.
• The model demonstrated adaptability to different data sampling rates and properties.

Conclusions:

• The BP-AR-HMM provides a robust statistical model for eye-movement classification, offering advantages over rule-based and fixed-category approaches.
• This method enables data-driven event parsing, uncovering a richer spectrum of eye movements.
• The BP-AR-HMM is effective for initial exploration and detailed analysis of gaze data characteristics.