Research Of Dynamic Link Prediction Algorithm Based On Link Importance

As a social network, the entity interaction in co-authorship network is that two authors write a paper together. An important problem in co-authorship network is to predict whether two authors will write a paper together in future. In data mining field, this problem represents link prediction. Current approaches generally solve it from two perspectives: graph topology feature computing or a semantic feature based classification, but all of this have two problems: one is that links generated by a paper are equivalent treated in computer network representation, but in fact, links between different authors play different role. Another is that none of proposed methods take into account the evolution history (time factor). Clearly, older links are less likely to be relevant for determining the future linkages than recent ones.According to the two problems mentioned above, a link importance based dynamic link prediction algorithm is proposed. A preprocessing will be implemented on DBLP dataset, including solution of linked sub-graph and filter of two rules, and then the dataset becomes small and less noisy. This process is useful for higher prediction accuracy and a low time complexity.After the preprocessing, a metric named link importance is proposed, and based on this metric, modification on classical topology feature is implemented. For semantic similarity computation, the TF-IDF algorithm used in vector space model is modified. According to link importance, per paper contribution of author is proposed to evaluate the suitable extent per paper on different authors. Based on the two factors, semantic similarity metric is proposed. Finally, time factor, which reflects the dynamic influence to generation of link, will be introduced to revise topology features and semantic similarity to generate the final definition.Based on these proposed methods, experiments are implemented on the DBLP dataset. The experiment results show that, rather than current approaches, link importance metric and time factor improve the prediction accuracy.