Spatial Saliency of Eye Movement Prediction Using Logistic Regression

We have always wondered why our gaze naturally darts to a person’s face in a photo, or fixates on a bright, moving object in a video. Our eyes are in constant motion, making rapid movements called saccades to points of interest, called fixations. For decades, scientists and engineers have been trying to unravel the secrets behind this process, because predicting eye movements is one of the most crucial factors in human-centred computing. It allows us to design better websites, create more engaging advertisements, develop diagnostic tools for neurological disorders, and even help AI understand human attention. “But how do you actually predict something as seemingly subjective as where a person will look”? In this paper, we explored one surprisingly effective and interpretable approach using Logistic Regression to model the spatial distribution of eye movements. We design a control system that tackles the problem from a sea of pixels to a point ofgaze where eye movement predictions is analysed and the system’s results is visuallised. The results shows increase in error from 0.01−0.6% from both logistic regression models (MDL1 and MDL2) and steady rise in learning from 80 − 100%. The initial result was a set of statistically significant coefficients (weights) assigned to each visual feature. These weights provide a clear hierarchy of what drives spatial attention as a high impact of local contrast. As expected, areas of high local contrast (e.g., sharp boundaries or text on the webpage) consistently received the highest positive weights, strongly predicting the spatial location of fixations. The aim of the paper demonstrates that Logistic Regression, when tailored to model spatial coordinates as data points, is an incredibly effective and interpretable tool for dissecting the patterns of human eye movement.