Preferential attachment based graph representation learning for Heterogeneous Networks

Abstract

A complex system can be modeled as a graph by considering entities as newlinenodes and interactions among entities as edges. Generally, real-world sys- newlinetems are complex and are modeled as heterogeneous networks having differ- newlineent types of nodes with multiple types of edges connecting them. It is well newlineknown that the preferential attachment (PA) mechanism drives link forma- newlinetion and evolution in the scale-free real-world networks. The idea of the PA newlinemechanism is that certain nodes are shown preference to form connections newlinecompared to the others. In real world social networks, the preferential fac- newlinetors may be popularity, influence, age etc. We believe that incorporating newlinesuch preferential factors into the algorithms would help in solving the prob- newlinelems on networks efficiently. In this thesis, we focus on modeling complex newlineinformation as a heterogeneous network and propose meaningful heteroge- newlineneous node and edge heuristics based on the preferential attachment mecha- newlinenism to solve the recommendation and link prediction problems. We restrict newlineour focus to the bibliographic information network, which is a prototypical newlineexample of a rich heterogeneous network. We conduct extensive experimen- newlinetation on a large DBLP v11 dataset, which contains 4,107,340 papers and newline36,624,464 citation relationships. For a few experiments, we also consider newlinesmaller DBLP v10, ACM, and Hep-Ph datasets for comparison purposes. newlineWe propose a preferential network evolution model (PNE) for directed net- newlineworks by extending Price s preferential attachment model for citation net- newlineworks, incorporating temporal factor such as recency into the model. We newlineprove that the model satisfies the power law and establish the same through newlinesimulations on the DBLP dataset. This also establishes the significance of newlinepreferential attachment in link formation and network evolution in citation newlinenetworks. The results of the PNE model match with the original DBLP network with high significance. We model the bibliographic information newlineas a heterogeneous network with paper, author, and jo