| Collaborative filtering aims to predict a person's preferences by examining the preferences of similar people. By necessity, many existing collaborative filtering algorithms rely on a coarse notion of similarity, a notion which assumes we can compare two people in terms of taste the same way we might compare them in terms of height or shoe size. Specifically, it assumes that if I like enough of what you do in a few specific areas, I am likely to make good recommendations for you in other areas. In fact, trust in this case is rarely implicit; more often we tend to trust recommendations from certain people in certain areas.;In this paper we introduce a notion of trust which reflects this quality. Rather than capturing taste information at the user level, we use tagging behavior to capture taste at the topic level. We find that doing so provides a significant improvement in the accuracy of recommendations without a commensurate loss in coverage. Also, this notion of trust naturally gives rise to networks which display interesting properties. We believe these networks can be exploited to further improve recommendation results, and investigate several possibilities which are inspired by recent research in network theory.;We test the theories above on a data set from CiteULike. This site allows researchers to save and tag articles of interest. Although the CiteULike data set has been used extensively in collaborative filtering research, we find that the data "as is" suffers from a significant spam problem. Part of the study below involves an investigation of how this spam changes the character of the data set. |
