Research On Financial Social Network Centrality Parallel Algorithms Based On Pregel

People and their relationships build up social networks.Social network analysis can apply to sociology,psychology,financial research or social media research.Network centrality measures the importance of individual in network,and reveals one's social relationship.Researchers in finance have worked on the relation between network centrality of chief executive officers and merger performance of their companies,and proved that there exists a positive relation.It indicated that social network analysis is valuable for financial research.The exponential increment of data brings the rapid expansion of social networks.Therefore,parallel algorithms of network centralities are designed gradually,to substitute the algorithms of serial computation.Apache Spark is a framework for distributed computation.Following Pregel parallel computation model,the Pregel API of Graph X library is for parallel graph computation.In this paper,n-Degree Centrality,k-Stress Centrality,k-Closeness Centrality and kHarmonic Centrality are proposed on the basis of Degree Centrality,Stress Centrality,Closeness Centrality and Harmonic Centrality respectively.The purpose of the latter two is to simplify the computation of Closeness Centrality and find the substitution suitable for financial social network analysis.Then,parallel algorithms for n-Degree Centrality,k-Stress Centrality and Closeness-like Centralities were designed based on the mechanism of Pregel model.Parallel algorithm for Closeness-like Centralities can apply to the centralities related to shortest-path distance.All these parallel algorithms were implemented through Apache Spark,and the iterations of graph computation would be completed by Pregel API.After that,we performed experiments on the centralities,the parallel algorithms,as well as financial social network.Most of the data come from Board Ex,a real financial social network database,and the rest are random networks generated from the network models.There are three partitions of experiments.Firstly,we recorded the time overhead that the parallel algorithms running on Spark cluster of different scale.Then we worked out the speedup of the parallel algorithms respectively and proved their availability and scalability.Secondly,we tested the similarity performance of k-Closeness Centrality and k-Harmonic Centrality with Closeness Centrality.We found that the variants are highly similar to Closeness Centrality not only in real social networks but also in random networks.For financial social networks of Board Ex,3-Harmonic Centrality comes with greater coefficients of variation and larger ranges,which means it has more remarkable distinction than Closeness Centrality.Therefore,it is possible to substitute Closeness Centrality with 3-Harmonic Centrality in further financial social network analysis.The last part is financial social network analysis.With Spearman's rank correlation coefficient,we conducted correlation analyses between each network centrality of different executives and the company value.The results showed that social relationships of executives have positive effects on the company values,among which the effects on market capitalizations are greater than that on revenues;2-Degree Centrality and 4-Degree Centrality are more generic in financial social network analysis with single centrality,as they could better reflect the positive effects of executives' social relationships on the company values;social relationships of chief technology officers and chief information officers have more notable effects on the company values than those of other executives in general.