The Link Prediction On Social Network

The link prediction on social networks is a new research direction of data mining. Socialnetworks represent social entities (such as individuals, families, networks, community groups,business organizations, national, etc.) and the relationships between them. Link prediction isfocused on explaining the hidden patterns and the impact of these relationships. Therelationships between the entities often contain a wealth of useful information, traditional datamining is often based on the attribute information of the entity, and link prediction of socialnetworks mine the relationship between the entities. Link prediction has a broad range ofapplications, such as in sociology, link prediction can be used to study the process of socialevolution; in bioinformatics, link prediction can be used to detect interactions betweenproteins; in the field of e-commerce, link prediction can be used to create a recommendationsystem; and in the security field, the link prediction can help to find the hidden terroristcriminals gang. Therefore, the study of the link prediction on social networks has animportant value.At present, the existing link prediction algorithm mostly predict the links based on thesimilarity between the two entities, and the social network topology information is a majorinformation source for designing the function of similarity between the entities. However, theexisting link prediction algorithm does not fully use the network topology information, suchas Common Neighbor algorithms only uses the number of common neighbors as thesimilarity function between entities, ignoring the interaction between the common neighbors;Katz algorithm uses the number of paths between nodes to evaluate the similarity of twonodes, treats the paths that have the same length equally, ignoring the different nodes in thepath. In addition, most of the link prediction algorithm is based on a static social network,ignoring the dynamic information of social networks. In this paper, these problems werestudied:Firstly, this paper proposes the concept of node guidance-force, node guidance-forcetakes full advantage of the network topology information, not only considers the differentnodes in the link prediction, but also takes into account the interaction between nodes.Secondly, as for the problem that the Common Neighbor algorithm can not distinguishthe contribution of a common neighbor on the link prediction, in this paper, nodesguidance-force is introduced into the link prediction algorithm, and the improved CNGFalgorithm is got, this algorithm gives the weight of each common neighbor. This paper alsogets the improved KatzGF algorithm to compensate for the deficiencies of Katz algorithm. This algorithm is based on global topology information. It not only gives the weights of thenodes on each path, but also distinguishes the contribution of the paths that have the samepath length on the link prediction effectively.Thirdly, considering the time attribute information of the node, putting the average degreeof nodes as the dynamic properties of the node, combining the node topology properties anddynamic properties, this paper gives the link prediction algorithm based on multi-attributes ofthe node.Finally, we verify the proposed link prediction algorithms on the DBLP data set. Theexperimental results show that the performance of CNGF algorithm and KatzGF algorithmare better than the Common Neighbor algorithm and Katz algorithm. And the performance ofthe link prediction algorithms based on multi-attributes of the node are better than the linkprediction algorithms based on single attribute of node.