Predicting answers in preference elicitation dialogs

The elicitation of user preferences represents an effective means to identify the relevant user preference requirements for a configuration task. One common method to determine the preferences of users is to survey users using dialogs that consist of (multiple-choice) questions with selectable answers. A major drawback of dialog-based preference elicitation is that many new users are not willing to answer a fairly large number of questions which is a prerequisite for the identification of user constraints. To that end, we have developed a novel similarity-based approach which aims to solve such ramp-up (cold-start) scenarios by automatically completing a set of remaining questions in a preference elicitation dialog given a small set of pre-answered questions. Our approach has been evaluated with two small real-world datasets. Initial evaluation results reveal that our approach is able to find the most probable answers a respondent is likely to give to a set of remaining questions. First insights of an evaluation also show that our approach can keep the number of initial questions at a very low level, meaning that only between 35% and 50% of questions have to be asked in most cases in order to predict the complete set of user requirements. The results also indicate that this level can be further reduced with increasing amounts of training data.