Exploition Of Location Information In Link Prediction Problem

The rapid development of information technology generates large amounts of network data. There have been great challenges on how to quickly and effectively find information from vast amounts of network data. In the field of social network analysis, link prediction is a key issue in the research of the network structure, which is not only to consider the content of the nodes’ properties, but also the relationship between nodes. Data mining from the location-based networks is an emerging research topic where the nodes have relationship properties as well as location properties.Currently, there are few researches on the link prediction in the location-based network and usually the prediction tasks of social links and location links were assumed to be independent. However, these two prediction tasks are strongly correlated in the real-world networks,. In addition, how to use amounts of unlabeled sample data to get more information effectively is also a problem in the real-world networks where the data has massive, sparse and dynamic characteristics. Finally, link prediction methods have emerged bottleneck and improving the traditional prediction methods and the accuracy rate is also a challenge.In this thesis, we first analyzed and summarized of the link prediction methods in the location-based network by exposing the challenges and raising the direction for solution. A semi-supervised link prediction method for location-based social network is then proposed, in which the location properties are extracted and the corelation between location and relationship are analyzed. The method introduces the semi-supervised learning method from machine learning to utilize the unlabled data more effectively. Finally, the experiments are conducted with the location-based Gowalla social network to compare with supervised learning method with and without location information and semi-supervised learning method without location information.The impact of the location on the form of friend relationship is also analyzed.Experimental results show that strong correlation exists between relationship and location in a location-based social network. The introduction of semi-supervised learning can improve link prediction accuracy, the addition of location information is also a certain help for the link prediction. These results open new directions for real world link recommendation systems on the link prediction methods and location-based social networks.