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

Use of Implicit User Feedback to Support Semantics-Based Personalized Document Recommendation 84 1. Introduction The advancement of information and network technology has facilitated the transfer, exchange and sharing of information on the Internet. However, for Internet users, the sheer volume of readily available information carries the risk of information overload. Previous research has suggested personalized recommender systems as a promising approach to simultaneously alleviate information overload and enhance user satisfaction (Liang, Lai, and Ku, 2006). To effectively provide appropriate, personalized information that appeals to their target users, many online stores have developed recommender systems that analyze important product attributes and customer characteristics (e.g., demographics, behaviors, or preferences) (Wang and Benbasat, 2007; Wei, Chen, Yang, and Yang, 2010; Xiao and Benbasat, 2007; Yang, Tang, Wong, and Wei, 2010). A typical recommender system makes personalized recommendations based on the target user’s profile and/or preferences. Amazon.com is famous for offering recommendation services by analyzing customers’ characteristics and preferences (e.g., purchased and rated items) across various products (Linden, Smith, and York, 2003). Along with sellers of physical products, many content providers have also equipped their websites with personalized content recommendation services for their readers. A good example of this approach is ScienceDirect, 1 a full-text scientific database offering journal articles and book chapters, which implemented a recommender system that automatically recommends articles that are related to the reader’s searches for specific articles. CiteULike 2 is a personalized digital library website that helps academics share, store and organize the academic papers they are reading, and provides article recommendations by analyzing users’ preference folders. Content-based filtering and collaborative filtering are two prevalent and promising approaches by which recommender systems can generate recommendations (Lee, Hu, Cheng, and Hsieh, 2012; Liang, Yang, Chen, and Ku, 2008). The content-based approach recommends products or services that might align with the focal user ’s known preferences, while the collaborative filtering approach makes recommendations in accordance with the choices of other users who share similar interests with the focal user. Although the content-based approach is considered to be more effective in locating textual 1 http://www.sciencedirect.com/ 2 http://www.citeulike.org/