Link Prediction Based On Dynamic Weighted Social Attribute Network

Social network websites provide a platform for users, where users can communicate with others, make friends and self-presentation. People’s communication will be more and more dependent on social networks. Links prediction for social network not only help researchers understand the evolution of the network, but also can provide friends recommendation services for users. Links prediction can increase the denseness of social network and users ’ stickiness to social website. Social network is a dynamic network, changing along with time. There are ties between entity nodes. The attribute information of entity nodes includes gender, age, interests and so on. It is necessary to take advantage of all possible data provided by social network in order to solve links prediction problems.This research treats links prediction as a binary classification problem, using machine learning method to train a cla ssifier for links prediction. Firstly, construct dynamic weighted social attribute network, which can take advantage of a variety of information provided by the social network. Secondly, extract temporal feature, network topological feature and attribute f eature from the network, including local node-based feature, global path-based feature. Use the samples’ feature vectors to train a classifier for links prediction.Most of the users on social network, only build friendships with little part of users. This paper designs a filter according to easy features and then filter large amount of negative samples from samples. Thus avoid computing unrelated nodes pair samples feature vectors, thereby reduce the time and space complexity of the prediction algorithm. There are a large number of implicit negative samples and little explicit positive samples in links prediction problem, which is called unbalanced data set. This paper keeps data balance by undersampling of negative samples, therefore can improve supervised algorithm performance.Entity nodes or links on social network are individual and different with each other. The difference between nodes or links is the theory basis for vertex weight and edge weight on social network. This paper designs several vertex weight and edge weight methods. The experiments prove that vertex weight should be inversely proportional to the nodes ’ centrality in local common neighbor feature so that get a better result. This research applies similar edge weight for links prediction, the more similar the two nodes are, the bigger edge weight value is. This paper also gets a better performance by using similar edge weight. Since social networks are dynamic, abstracting temporal feature from dynamic network can get better results than no temporal feature.