Recently, most of prediction models employ basic attributes of users or social relationships as the original features to optimize prediction results. Nevertheless, in the real systems, it’s highly difficult to acquire attributes and social information of users, which always lack of sufficient accuracy, so that user behavior couldn’t be expressed accurately as well as prediction models couldn’t be applied widely. Additionally, almost the whole models employing user features focus on measurement of interests of user itself, while ignoring the feature of the change of user interests. Based on the above considerations, we propose a collaborative prediction model based on transfer features of user interests. The model constructs network evolution of user behavior and transfer features of user interests according to the sequences of consistent user behavior, of which the former is used to compute similarity of users to retrieve nearest neighbor sets while the latter is applied to filter candidate sets. With the application of a news browsing log, as well as an internet protocol television access log and a microsoft server log, experiments indicate the effectiveness and practicability of our prediction model.
