Research On Key Technologies For Recommending User-Generated Content In Social Media

With the development and prevalence of Internet and mobile devices, World Wide Web has shifted from traditional media where users can only simply read the already-created content in a passive way, to social media where users can create and share content in an active way. Although this shift greatly facilitates the creation and sharing of in-formation, it also exacerbates the information overload problem. Due to some new properties of social media such as the time-effectiveness of information, the person-alization of information need, and the arbitrariness of information consumption, tra-ditional information filtering technologies such as information retrieval are becoming less effective in dealing with the information overload problem in social media. Dif-ferent from traditional technologies, recommender system aims to automatically ana-lyze users’potential and personalized information need from social media data, which makes it more suitable to deal with the information overload problem in social me-dia. However, existing recommender systems are facing great challenges in terms of how to comprehensively manage different social media data, how to model the com-plex user relationships, how to deal with the implicit user feedback, how to handle the time-effectiveness of user-generated content, etc. To address these challenges, we first propose a recommendation framework, and then put our focus on three key tech-niques with regard to user relationships, user feedback, and user-generated content, respectively. Based on these techniques, we design and implement a prototype recom-mender system, and further apply the system on Sina Weibo for recommending Weibo messages. In particular, we make the following efforts in this dissertation:1. We propose a recommendation framework for recommending user-generated content in social media. By analyzing the requirements of the user-generated content recommendation problem, the proposed framework aims to manage dif-ferent types of social media data including user relationships, user feedback, and user-generated content in a loosely coupled and flexible manner.2. We propose an approach for user relationship inference. Using a variety of soci-ological theories to depict the complex user relationships, the proposed approach aims at inferring some hidden user relationships from existing ones. Experimen-tal evaluations on several real data sets demonstrate that the proposed approach significantly outperforms several benchmark methods in terms of inference ac-curacy. In addition, it enjoys linear time complexity in the pre-computational stage and constant time complexity in the online response stage.3. We propose an approach for one-class user feedback inference. Due to the fact that user feedback is usually implicitly expressed in social media, the proposed approach aims to handle such implicit feedback by some special treatments, and further incorporates side information such as user relationships to better capture users’information need. Experimental evaluations on real data sets show that the proposed approach achieves significant accuracy improvement over several benchmark methods while enjoying linear pre-computation scalability and con-stant online response.4. We propose an approach for content-based user preference prediction. To handle the time-effectiveness of user-generated content, the proposed approach provides incremental algorithms to handle the case when user-generated content contin-uously arrive in a dynamic stream-like way. It also contains a family of variant algorithms to cover the choices when a linear or a non-linear model is selected or whether the associations between user-generated content are considered. Ex-perimental evaluations on several real data sets show that the proposed approach achieves a good balance between the prediction quality and the efficiency.5. We design and implement a prototype recommender system for recommending user-generated content in social media, and further apply the system on Sina Wei-bo for recommending Weibo messages. The implemented system is deployed and served as a platform. Based on the platform, we build an example appli-cation on Sina Weibo to show the feasibility of the proposed recommendation techniques.