Research On Pervasive Link Prediction Model For Multiple Types In Heterogeneous Academic Networks

With the rapid development of scientific research, the number of academic resources has increased explosively. Faced with the huge academic resources, researchers hope to be able to quickly and accurately find the relevant papers, authors or journals that they can contribute and other academic entities. In the face of these challenges,more and more researchers have studied the academic networks from different perspectives. It is of great significance to construct homogeneous or heterogeneous academic networks and forecast the missing or future links by link prediction methods, so as to achieve the objectives of entity recommendation, relation prediction and network analysis.The current link prediction methods for heterogeneous academic networks can only predict the single or limited types of relationship. The constructed link prediction model has poor generalizability and the accuracy of link prediction needs to be improved. Aiming at the problems above, this paper proposes a pervasive link prediction method for heterogeneous academic networks. This method constructs a heterogeneous academic network which includes the common types of academic entities and the types of relations between entities. This paper proposes an automatic meta-path extraction method to mine the deep characteristics in the network, and proposes two improved methods which consider time and the content relevance between entities respectively to measure the similarities between meta-paths. And finally the multi-relation universal logistic regression model which can be used to learn characteristic parameter is chosen to construct the final link prediction model.We use Microsoft Academic data as the experiment dataset to construct heterogeneous academic network, and make automatic link prediction for a variety of relationship types with real meaning in the network based on the pervasive link model proposed in this paper. The experimental results show that the link prediction model can be used to realize the automatic prediction of multiple link relations, and the accuracy of link prediction is improved obviously. Meanwhile, the time extension experiments prove that the model has good applicability over time.