Research On Technologies And Methods Of Social Tag Recommendation

The soul of Web 2.0 is allowing users to create content. As an application of Web 2.0, social tagging system has become a most popular application due to its lower barrier and easier usability. Tags are borned from social tagging systems and have many merits such as organizing, sharing, searching and finding information. However, there are still some drawbacks with tags, such as:1) tags are sparse yet in many social tagging systems.2) Users may lable tags arbitrarily for a resouce.3) Tags have a low level of utility.4) There are too many spams yet. All these drawbacks abate their benefits in many applications abovementioned. Hence, social tag recommender system calls a lot of attention from academic and industry circles recently. This dissertation studies these questions as follows.1) Proposed word feature-based and latent topic-based tag recommendation algorithms. Due to the different granularity of the content of a given resource, we can present the content with word in fine-grained; also present the content with latent topic in coarse-grained. According to different presentation, different tag recommendation algorithms are proposed. We use language model to model the descriptions and tags of a given resource for fine-grained tag recommendation, and use Latent Dirichlet Allocation (LDA) to model the descriptions and tags of it for coarse-grained tag recommendation. Experiments showed that the most appropriate feature granularity for tag recommendation is word-grained. If we hybrid a singe model with topic featured-based model, a better result will be achieved. Meanwhile, if too many entities are introduced into modeling, it will cause much noise and impair the final recommendation.2) Proposed a topic sensitive tag ranking algorithm (TSTR). Usually, some resources labeled with a tag in some topics are so dominated that other resources labeled with the same tag are not found out. This makes a negative effect on information retrieval and information utility in precision and recall. This paper uses LDA model to extract all topics in tag space. Following the co-occurrence relationship and the distribution of topics, we construct a tag hypergraph based on topics and compute the authority score of a tag in line with the distribution of topics. Then, we use the authority score for tag recommendation. Experiments showed that the effection of the proposed algorithm of tag-based topic recommendation is great better than other traditonal algorithms.3) Proposed a tag recommendation model based on users'motivation orientation. In order to enhance the usability and viscousness of social tagging system, speed up the progress of stabilizing and semantic emergence, we use five metric indices to measure the motivation orientation of users, and then group users into describers and categorizers after analyzing the tag space deeply. After validating the five metric indices, a tag recommendation model based on users'motivation orientation was proposed. When tagging, we present a user and a resource using the metric indices vector. If the motivation of a resource matches the motivation of a user, we aggregate all tags of the resource into a set for computing the authority score and relatedness to the resource which the user wants to label. The top k tags will be selected for the resource according to the product of the authority score. We conduct experiments on two different datasets, and the results showed that the performance of the proposed model is better than the baseline models.