What surprises does your past have for you?. (November 2017)

• Record Type: Journal Article
• Authors: Mourão, Fernando; Rocha, Leonardo; Araújo, Camila; Meira, Wagner; Konstan, Joseph
• Is Part Of: Information systems. Volume 71(2017)
• Page Start: 137
• Page End: 151
• Publication Date: 2017-11
• Subjects: Recommendation -- User modeling -- Unexpectedness; Database management -- Periodicals; Electronic data processing -- Periodicals; Bases de données -- Gestion -- Périodiques; Informatique -- Périodiques; Database management; Electronic data processing; Periodicals; 005.7
• DOI: 10.1016/j.is.2017.08.001
• Languages: English
• ISSNs: 0306-4379

Abstract

Highlights: Proposal of a methodology of usefulness quantification for recommending known items. Proposal of distinct strategies for identifying unexpected known items in real domains. Offline and Online experiments that demonstrate the usefulness of this kind of recommendation. Abstract: Although the current Recommender Systems (RSs) focus on discovering unknown items for users in several domains, users may be particularly interested in consuming items in which they are already familiar. As a result, this study aims to uncover subsets of known items that are useful for recommendations in the present. The main argument highlighted in this study is that past consumption is a rich source of relevant recommendations neglected or underexploited by current RSs. Thus, we propose a methodology to quantify the effectiveness of recommending known items in real domains. Afterwards, we proposed distinct heuristics to search the consumption history of each user items unexpected to be consumed, but potentially relevant. Such heuristics exploit time-related, context-related, and relevance-related information; as well as a combination of these three types of information. Assessments on real collections allowed us to verify the applicability of our methodology. Furthermore, offline evaluations demonstrated that past relevance, consumption recency, and associations with currently consumed items are useful information to model reconsumption. Finally, through a user study with members of Movielens, we verified that users are willing to reconsume some known items and recognized the value in this type of recommendation. Therefore, by exploiting the long history of each user, we are able to match a piece of the user's preference neglected by current RSs, hence, improving the user's satisfaction.