Prediction of social bookmarking based on a behavior transition model

We propose an algorithm to predict users' future bookmarking using social bookmarking data. It is a problem that primitive collaborative filtering cannot exactly catch users' preferences in social bookmarkings containing enormous items (URLs) because in many cases user's adoption data is sparse. There can be various influences on bookmarking such as effects from the environment and changes in user preference. We use temporal sequence among the bookmarking-users to represent word-of-mouth and among the bookmarked-URLs to represent user's interest, and model each sequential order as a continuous-time Markov chain. This idea comes from diffusion of innovation theory. A transition probability from a state (user/URL) to another state is defined by the transition rate calculated from the time taken for the transition. We predicted user's preferences through a combination of estimating the most likely transition between users using URLs as input and between URLs using users as input. We conducted evaluation experiments with a social bookmarking service in Japan called Hatena bookmark. The proposed algorithm predicts users' preferences with higher accuracy than collaborative filtering or simple transition models based on either user or URL.