< Back to previous page

Publication

Mixed-initiative Recommender Systems: Towards a Next Generation of Recommender Systems through User Involvement

Book - Dissertation

Recommender systems have been ubiquitous in our daily lives. However, most existing recommenders do not explain the logic or justification for the resulting recommendations. This black box nature may hinder users from accepting the recommendation due to low user understanding and trust. To address this issue, several researchers have proposed an approach that leverages interactive visualization techniques to improve transparency and controllability of recommenders. Whereas this approach has been evaluated in several user studies, the roles of transparency and user control have not been fully investigated yet. Moreover, most recommenders often offer different users the same control components. This "one-size-fits-all" approach does not meet the diverse user needs and contexts of use. This dissertation intends to help system designers design mixed-initiative recommenders that optimize the user experience through personalized user interaction. This dissertation aims to address four research questions: 1) RQ1: How do transparency and user control influence the user perception of recommendations?; 2) RQ2: How do personal characteristics influence user perceptions of recommenders with user control and visualizations?; 3) RQ3: How do contextual characteristics influence user perceptions of recommenders with user control ?; and 4) RQ4: How do natural user interfaces facilitate user control of the recommender system?. The exploration of these research questions include the design, implementation, and evaluation of the following research prototypes: 1) a user interface to support transparency and user control for product recommendation, 2) a user interface to support different levels of user control for music recommendation, 3) a user interface to support user control on context characteristics for music recommendation, 4) a visual search based user interface for product recommendations, and 5) a conversational user interface for critiquing-based music recommender. Also, we investigate how personal characteristics and contextual characteristics influence the impact of user interface design on recommenders. The results highlight the benefits of transparency and user control and the influence of personal characteristics and contextual characteristics on the design of user interfaces. In particular, we found that 1) combining transparency and user control improves user perception to the greatest extent; 2) enabling all levels of user control together leads to the highest acceptance without increasing cognitive load, but more sophisticated visualization does not lead to higher perceived diversity; 3) musical sophistication and desire for control significantly influence user perception and user interaction; and 4) mood significantly influences perceived quality and diversity and activity influences user behavior of controlling systems. Besides, the results of evaluating a conversational interface and a visual search user interface also shed light on employing natural user interfaces to enhance user control in recommenders. In the end, we also present design guidelines for implementing different components of mixed-initiative recommenders investigated in this dissertation.
Publication year:2019
Accessibility:Closed