Common structure and properties of filtering systems Junichi Iijima * , Sho Ho Graduate School of Decision Science and Technology, Tokyo Institute of Technology, Tokyo, Japan Available online 11 December 2006 Abstract Recommendation systems have been studied actively since the 1990s. Generally, recommendation systems choose one or more can- didates from a set of candidates through a filtering process. Methods of filtering can be divided into two categories: collaborative filter- ing, in which candidates are chosen based on choices of other persons whose interests or tastes are similar, and content-based filtering, in which items are chosen based on the profile or action history of the recommendee. However, these methods share the same structure in the sense that both of them recommend items based on relevance degrees of items and references, as well as relevance degrees between the recommendee and each reference. Most discussions about recommendation systems focus on the methods of choosing recommended candidates; few focus on foundational concepts of recommendation conditions that systems must satisfy, and problems that current sys- tems have compared with these conditions. In this paper, recommendation systems are reconsidered from the viewpoint of multi-criteria decision making. Conventional filtering methods (e.g., collaborative filtering and content-based filtering) are formulated as linear weighted sum type recommendation systems. Several properties of linear weighted sum type recommendation systems are identified and formulated from the viewpoint of voting. Ó 2007 Elsevier B.V. All rights reserved. Keywords: Social choice theory; Collaborative filtering; Content-based filtering; Recommendation system 1. Introduction As an information and communication technology (ICT) application that meets the growing needs of ‘‘person- alization’’ in our advanced information society, recommen- dation systems have been studied actively since the 1990s. ICT-enabled recommendation systems (e.g., online shop- ping systems that recommend products based on customer profile and history of customer actions; enterprise knowl- edge portals that send necessary information in a timely manner to each employee according to specialty and posi- tion) are infiltrating various aspects of our life. On the other hand, due to the rapid growth of ICT, an enormous amount of information that exceeds the capability of human information processing is now being distributed via various networks. To handle the flood of information, recommendation systems that effectively collect and choose information based on objectives and preferences of users are becoming indispensable. Early studies of ICT-enabled information recommenda- tion systems include Tapestry [1], GroupLens [2], and Fab [3]. The term ‘‘collaborative filtering’’ was first used in 1992 by Goldberg et al. in their paper that introduced an infor- mation distribution system named Tapestry. In Tapestry, users can set rules such as ‘‘if Joe and Bill receive a mes- sage, then I would like to receive that message, too’’ and filter messages based on the rules. The collaborative filter- ing in Tapestry was conducted semi-automatically. In the late 1990s, GroupLens, which conducts collaborative filter- ing automatically, was developed. GroupLens is a system for collaborative filtering of online news in which the rela- tive degree of users is calculated based on their rating of articles, and articles are recommended based on the rate given by highly relative users. Meanwhile, hybrid recom- mendation systems that integrate content-based filtering and collaborative filtering have also been developed. Fab, which recommends WWW pages to users, is a typical hybrid recommendation system. In the Fab system, users 1567-4223/$ - see front matter Ó 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.elerap.2006.11.002 * Corresponding author. E-mail address: iijima@me.titech.ac.jp (J. Iijima). www.elsevier.com/locate/ecra Electronic Commerce Research and Applications 6 (2007) 139–145