Temporal Correlation between Social Tags and Emerging Long-Term Trend Detection

Social annotation has become a popular manner for web users to manage and share their information and interests. While users' interests vary with time, tag correlation also changes from users' perspectives. In this work, we explore four methods for estimating temporal correlation between social tags and detect if a long term trend emerges from the history of temporal correlation between two tags. Three types of trends are specified: steadily shifting, stabilizing, and cyclic. To compare the results of the four estimation methods, an indirect evaluation is realized by applying detected trends to tag recommendation. Introduction With the growth of Web 2.0, social annotation services such as del.icio.us, YouTube, and Flickr have been important manners of organizing information on the web (Hammond 2005). These Web 2.0 sites provide users with the functionality of sharing interesting and useful information with friends and even with the public, in a malleable, convenient and ease-to-use fashion. User-generated metadata in such services are often referred to as tags. Since tags reflect users’ perception and interpretation of target resource (Li, Guo, and Zhao 2008), here we consider tags as interesting concepts for users. The rapid popularization of social tagging has attracted considerable works for analysis or utilization of such rich metadata (Bao et al. 2007; Golder and Huberman 2006; Halpin et al. 2007; Wu et al. 2006; ). In these applications, estimation of semantic correlation between two tags (or, concepts) is fundamental and indispensable. While users’ interest may shift as time goes on, correlations between various concepts may shift from users’ perspectives. The following example clarifies this idea. It concerns a hot topic in del.icio.us “website design programming”. Intuitively users may be interested in different programming languages in various periods. As a result, the semantic correlation Copyright © 2010, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. between “website design” and a specific programming language (e.g., “PHP”) vary with time. As far as our knowledge, the time factor is not concerned in previous researches on social tagging. In this work, we attempt to estimate temporal correlation between two concepts and explore if there is any long-term trend emerging from the history of considered temporal correlation. An emerging trend indicates how user interest varies with time. There are three main issues to be addressed for this novel problem: 1) How to model the history of temporal correlation between two tags in an efficient manner? 2) What types of long-term trends are reasonable to be detected? 3) How to detect the emerging trend of a specified type? The fundamental assumption of this work is: appearance of an annotation reflects the condition that at the time when the annotation was generated, the annotator was interested in the concepts she assigned. Temporal correlation between two concepts is thus estimated by using the co-occurrence information. We model the evolutional history of concept correlation over a long period by estimating temporal correlation between two concepts in each sliding time frame in the period. We then specified three types of emerging trends in the history of temporal correlation: steadily shifting, stabilizing, and cyclic. For the task of trend detection, we employed typical regression models with various predictor functions.