Online Learning of Visualization Preferences through Dueling Bandits for Enhancing Visualization Recommendations

Abstract

A visualization recommender supports the user through automatic visualization generation. While previous contributions primarily concentrated on integrating visualization design knowledge either explicitly or implicitly, they mostly do not consider the user's individual preferences. In order to close this gap we explore online learning of visualization preferences through dueling bandits. Additionally, we consider this challenge from a usability perspective. Through a user study (N = 15), we empirically evaluate not only the bandit's performance in terms of both effectively learning preferences and properly predicting visualizations (satisfaction regarding the last prediction: μ = 85%), but also the participants' effort with respect to the learning procedure (e.g., NASA-TLX = 24:26). While our findings affirm the applicability of dueling bandits, they further provide insights on both the needed training time in order to achieve a usability-aligned procedure and the generalizability of the learned preferences. Finally, we point out a potential integration into a recommender system.