User Interest Discovery And Personalized Information Recommendation For Online Forum

As a typical Web2.0application, online forum attracts millions of users to contribute a large number of valuable contents. How to extract useful knowledge from online forum remains a big challenge. Especially, the traditional information retrieval techniques can not deal with this task efficiently for the lack of real-time response, effective ranking and the capability of handing non-keyword search. Active personalized recommendation will be a strong way to capture information in online forum.This thesis mainly focuses on personalized recommendation in online forum. Com-pared to those traditional recommendation systems, there exist two challenges in person-alized recommendation.·It’s challenging to understand user’s interests for online forum information searching. Forum contents are short and consist of various user-generated contents(UGC). In addition, social networking and user-generated contents are mixed up and mutually reinforced. Because of the absence of explicit rating scores for recommended objects, it is also hard to measure similarity between users and recommended objects.·It’s challenging to find efficient match algorithms between recommended objects and user interests. The recommended objects in forum have many styles, including real-time objects and non-real-time objects. The former has a strong timeliness, such as hot topics, events, and so on. The latter has a long lifetime and stays stable, such as field experts. Personalized recommendation is triggered by recommended objects, which is very different from the traditional recommendation pattern driven by target user.In order to handle the above two challenges, this thesis proposes a framework of person-alized recommendation in online forum, which includes three models:how to find user interest profile, how to find objects for recommendation and how to match between user interests and recommended objects. The major contributions are as follows,1. Based on analyzing the features of forum data and the existing techniques of user interest discovery, we propose two graph models for active users and in-active users respectively. In both graph models, Random Walk with Restart(RWR) is used for capturing user interests. A big challenge for user interest discovery is that social networks and contents are mixed up and have a mutual influence and reinforcement. Fortunately, the tripartite graph model for active users and the social context graph model for in-active user proposed in the thesis can integrate social networks with contents well. The direct and indirect relationships between social networks and contents can also be obtained by Random Walk with Restart algorithm in these two graph models.2. After reviewing existing recommendation algorithms, we propose a novel reverse rank-ing query, top-k reverse ranking query, to match user interests and recommendation objects. A framework of responsing top-k reverse ranking query is presented. The top-k reverse ranking query framework determines the upper bound and lower bound of rank score of query point based on the geometry properties of top-k query and the MBR(Minimum bounding rectangle) property of R*-tree, which can quickly prune the candidates of weight vectors.3. Proposing a solution to discover a kind of major recommended objects in online forum, namely filed experts. It is very helpful for users to find experts in the corresponding fields, since experts’opinion will influence the recommended objects in online forum. In this thesis, we study how to find experts and their profiles in online forum. First, we use a tripartite graph model to capture both contents and structure information in online forum. Based on this graph model, a single random walk with restart procedure is applied to evaluate the relevance between a user and a term. Subsequently, a star-based optimization method is proposed to accelerate the RWR computation over tripartite graphs. Analysis shows that this method can significantly reduce the online computation cost since it reduces the size of transition matrix.As a conclusion, this thesis carries out a detailed study on discovering user interests, matching between user interests and recommended objects and discovering of experts. On this basis, the personalized information recommendation framework is proposed for online forum. All related work is based on the comprehensive survey and analysis on existing theories and techniques. The theoretical analysis and extensive experiments show that the proposed approaches perform well to discover user interests and recommended items, and match between user interests and recommended objects.