Recommender systems are useful information systems that can solve information overload problem and provide personalized services. Their “push information” mechanism complements the “pull information” mechanism of information retrieval. The main values of recommender systems lie in: 1) to help users discover and extend their interests; 2) to help content providers send content to interested users in a friendly way. The general idea of recommender system design is to predict users’ interest in items by applying data mining, machine learning, and statistics etc. to process explicit or implicit user-item interaction data and the features of users and items. The evaluation of recommender systems mainly includes these aspects such as accuracy, coverage, diversity, novelty, and serendipity. There are different metrics for each aspect. In recent years, many social media tools have emerged, which provides new challenges and chances for recommender systems in terms of data and recommendation system type. In this dissertation, we have studied recommender systems in social media regarding the following three aspects: 1) how to use user-tag-item information to solve data sparsity problem for social tagging systems; 2) how to improve the diversity and novelty of recommendation results by exploiting tag information in social tagging systems; 3) to study friend recommendation and information diffusion in social networks by spectral methods. The contributions of this dissertation are as follows. 1. A random walk model for item recommendation in social tagging systems In order to alleviate data sparsity problem in the context of social tagging systems, we proposed a random walk model for item recommendation which can effectively exploit the transitive associations among users, tags, and items. Using these associations can make the measure of the associations between entities such as user-item, user-user, and item-item etc. more accurate. Therefore, we designed a PageRank-like algorithm which can capture the transitive associations and obtain personalized item ranking lists through multi-step random walks on item graph and user graph. Empirical evaluation on three real-world datasets demonstrates that our approach can effectively alleviate the sparsity problem and improve the quality of item recommendation. 2. An approach to improve the diversity and novelty of recommendation by exploiting tag information As to social tagging data, we proposed a novel item recommen...
修改评论