Research On Link Prediction In Networks With Nodes Attributes

In the scientific research process, every discipline will inevitably encounter the problem of prediction, link prediction is an important research area in the field of data mining. In the real world, a large complex system in the nature, or an artificial system which can help us analysis and solve the problem, can be described and characterized as a complex network. In these systems, the relationships between the individuals are usually complex, In the evolution of such networks, due to the limitations of space and time or the experimental condition, missing or redundant links may occur in the network, and some potential links may not be discovered. Furthermore, the structure of a network are dynamically changing over time, Therefore, links may delete or added to the network from time to time. We need to predict the missing and potential links in the network, this is the problem of link prediction in complex networks. Link prediction helps us analysis and understand the topological structure of the networks.In the early stage, methods of link prediction are mainly based on the Markov chain and the machine learning. In recent researches, people realize that topology structure of the network clearly shows the structural similarity between nodes of network, and the similarity often reveal whether a link exists between them, so scholars have put forward various index to depict the structure similarity, such as CN, AA and Katz, these indexes have shown better prediction effect. In some application problems, nodes in the network have attributes which contain rich information about the contents and characteristics of the individuals of nodes. Those information affects the existence of the links in the network. If we combine the structure information and attribute information together, we can greatly improve the prediction accuracy of link prediction algorithm.Nowadays, how to organically integrate the structure information and attribute information is a hot topic in network analysis. In this work we study the problem of link prediction in networks with nodes attributes. We present approaches for integrating the the structure and attribute infonnation, and efficient algorithms for predicting links in networks with nodes attributes to improve the precision of prediction.The research work and main contributions of this paper are as follows:(1) We propose a link prediction algorithm based on similarity propagation. In the algorithm the attribute similarity of nodes will be attached to each edge as a weight. Based on the shortest path algorithm, we get the reachability of all nodes in the network. According to the capacity of this reachability and the transfer rule, similarity between every pair of nodes will propagate in the whole network, and update their values in the process of propagation. We can get the possibility of the link between every pair of nodes after the process of propagation convergence. The experimental results on the relevant data sets show that the prediction accuracy of our proposed algorithm is higher than the other algorithms.(2) We propose a link prediction algorithm based on the selection of parameters. Considering that there is a parameter needs to set in the Katz index computation, we optimize the value of such parameter according to the attributes on the nodes and get a proper value of the parameter so as to integrate the structure and the attribute information efficiently. Our experimental results show that our proposed algorithm can obtain high quality prediction results.(3) We propose a prediction algorithm based on space mapping. We map the attribute information and structure information of network into the same space, aiming to calculating the similarity in the new space. In the mapping process, we use of the method of matrix factorization. After completion of mapping, we use the method of matrix alignment to make the structural similarity and attribute similarity become consistent. Then we can obtain the score matrix. Experimental results on the real data sets show that the prediction accuracy of our algorithm is higher than the other similar algorithms.