Detecting unfair recommendations in trust-based pervasive environments

Abstract In pervasive/ubiquitous computing environments, interacting users may evaluate their respective trustworthiness by using historical data coming from their past interactions. Nevertheless, when two users are at the first interaction, they have no historical data involving their own activities to be analyzed, and then use information (recommender-data) provided by other users (recommenders) who, in the past, have had interactions with one of the involved parties. Although this approach has proven to be effective, it might fail if dishonest recommenders provide unfair recommender-data. Indeed, such unfair data may lead to skewed evaluations, and therefore either increase the trustworthiness of a malicious user or reduce the one of a honest user in a fraudulent way. In this work, we propose an algorithm for identifying false recommender-data. Our attention is explicitly focused on recommender-data rather than recommenders. This because some recommenders could provide recommender-data containing only a limited (but specific) set of altered information. This is used by dishonest recommenders as a tactic to avoid being discovered. The proposed algorithm uses association rules to express a confidence-based measure (reputation rank), which is used as a reliability ranking of the recommender-data. The resulting approach has been compared with other existing ones in this field, resulting more accurate in finding out unfair recommender-data sent by dishonest recommenders.