臺大管理論叢 NTU Management Review VOL.28 NO.3

87 NTU Management Review Vol. 28 No. 3 Dec. 2018 advantages over keyword-based systems. Their approach adopts the spreading activation model (SAM) to expand the array of concepts that are semantically relevant to the focal user’s preferred documents. Based on the semantic linking technique, Xu, Wei, Luo, Liu, Mei, Hu, and Chen (2015) developed “Knowle,” an online news management system. The Knowle system comprises three layers (concepts, resources and events), and is devoted to organizing and discovering online news. The core elements to the Knowle system are the news events that link to each other through their semantic relations on the Web (Xu et al., 2015). Cantador, Castells, and Bellog (2011) proposed a three-fold knowledge representation based on a semantic preference spreading mechanism to overcome two main challenges of recommender systems: cold-start and sparsity. 2.2 Implicit Elicitation A user-centered recommender system is not only equipped with an effective recommendation algorithm, but also takes the user experience into account (McNee, Riedl, and Konstan, 2006; Pu, Chen, and Hu, 2012). Hence, enhancing the user experience is an important user interface design issue for recommender systems. From the information system architecture perspective, a recommender system is composed of three units: input, process, and output. The function of the input unit is to acquire and record users’ profiles, browsing behaviors, and preferences. These personal data are manipulated by the process unit to predict a particular user’s preference for items that this user has never seen or bought, and to make appropriate personalized recommendations. The output unit presents the recommendations and helps users make decisions. The purpose of our study is to elicit implicit user feedback to enhance the user’s experience on a semantic- based recommender system. Thus, we focus our discussion on the user preference elicitation mechanism. The techniques of user preference elicitation can be classified as either explicit or implicit (Pommeranz, Broekens, Wiggers, Brinkman, and Jonker, 2012). Explicit elicitation requires the user to explicitly evaluate items in order to build the user profile, while implicit elicitation implicitly observes and analyzes the user’s activities to generate the user profile. Explicit elicitation requires more user effort in the form of active responses, e.g., completing a questionnaire, providing opinions such as ratings or comments, or indicating the degree to which something is liked or disliked. Besides, privacy concerns make users hesitant to provide personal information (Knijnenburg and Kobsa, 2013; Lee and Cranage, 2011). In contrast, implicit feedback requires no user