Abstract
In a decision-making scenario, one group member may need, independently from the other members, to choose an item, e.g., a restaurant, that will be experienced by the group. In a prior research on restaurant recommendation, we have identified two primary tasks of such an organizer of a lunch/dinner event, who is in charge of selecting a proper restaurant for the group: anticipating the other group members' preferences, to properly enter these preferences in the recommender system, and reconciling incompatible preferences, if they arise in the elicitation phase. To support the first task, we augmented a group recommender system with a machine learning model that predicts group members' food preferences, about dishes that can be consumed in the restaurant, based on other available information about the members (demographics and preferred cuisine). However, in a user study, we found that supporting functionality to be not effective in improving the quality of the choices made by the organizer. In this paper, we investigate the causes of this poor performance. We analyze the possibility that the poor performance may relate to: the deployed ML models used for food preference prediction, the amount of data about food preferences used for training the ML model, and the complexity of the preference anticipation task. The results of our experiments suggest that neither the ML model nor the scarcity of the preference data is responsible for the observed poor value of the implemented preference recalling function. While a user study seems to confirm the impossibility of accurately predicting food preferences from the considered user's characteristics (demographics and preferred cuisine).