Prediction for Occurrence of Character Identification Difficulty during Web Browsing by Apprenticeship Learning and Gaze Data
Matsushita Yutaka,
Aoki Shusei
Aiming to build a web-browsing assistance system that enlarges characters when users have difficulty
identifying them and fix their gaze at them, this study proposes a method for predicting the occurrence of
character identification difficulty during web browsing based on eye-tracking data. A criterion is needed
to determine when characters should be enlarged. However, establishing such a criterion is highly
challenging because the fixation duration associated with identification difficulty is not necessarily
longer than that of resulting from other factors. Therefore, the criterion is calculated by introducing the
user's saccadic velocity and amplitude as external parameters and classifying their combinations. In this
process, the study employs apprenticeship learning (a form of inverse reinforcement learning), thereby
eliminating the need to evaluate rewards. Moreover, a new algorithm, called “Recurrent Two-Step
Linear Programming Apprenticeship Learning (RTLPAL),” is developed based on linear programming
to reduce processing time and improve accuracy. This algorithm is characterized by the recursive
repetition of two processes: deriving the optimal policy for the apprentice and generating an improved
expert (reward). When applied to eye-tracking data from 10 participants during web browsing, RTLPAL
improves accuracy metrics and reduces processing time to approximately 1/20th of that by SARSA.
