A prediction-oriented optimal design for visualisation recommender systems

A good visualisation method can greatly enhance human-machine collaboration in target contexts. To aid the optimal selection of visualisations for users, visualisation recommender systems have been developed to provide the right visualisation method to the right person given specific contexts. A visualisation recommender system often relies on a user study to collect data and conduct analysis to provide personalised recommendations. However, a user study without employing an effective experimental design is typically expensive in terms of time and cost. In this work, we propose a prediction-oriented optimal design to determine the user-task allocation in the user study for the recommendation of visualisation methods. The proposed optimal design will not only encourage the learning of the similarity embedded in the recommendation responses (i.e., users' preference), but also improve the modelling accuracy of the similarities captured by the covariates of contexts (i.e., task attributes). A simulation study and a real-data case study are used to evaluate the proposed optimal design.