Abstract
Most recommendation systems require some form of user feedback such as ratings in order to make personalized propositions of items. Typically ratings are unidimensional in the sense of consisting of a scalar value that represents the user’s appreciation for the rated item. Multi-criteria ratings allow users to express more differentiated opinions by allowing separate ratings for different aspects or dimensions of an item. Recent approaches of multi-criteria recommender systems are able to exploit this multifaceted user feedback and make personalized propositions that are more accurate than recommendations based on unidimensional rating data. However, most proposed multi-criteria recommendation algorithms simply exploit the fact that a richer feature space allows building more accurate predictive models without considering the semantics and available domain expertise. This paper contributes on the latter aspects by analyzing multi-criteria ratings from the major etourism platform, TripAdvisor, and structuring raters’ overall satisfaction with the help of a Penalty-Reward Contrast analysis. We identify that several a-priori user segments significantly differ in the way overall satisfaction can be explained by multi-criteria rating dimensions. This finding has implications for practical algorithm development that needs to consider different user segments.