Research On Motif-based Social Network Users Ranking Algorithm

Nowadays,the online social network has become an important carrier and channel for our daily information transmission.The research on the social network has practical application value for the development of human society and commercial promotion.One of the most important issues in social network application analytics is ranking users.The existing social network user ranking algorithms are mainly divided into the centralitybased method,hyperlink-based topic search algorithm,and PageRank-based algorithm.Traditional ranking algorithms,whether weighted or unweighted,only use edge-based relations.Some ranking algorithms take into account the high-order structure that nodes participate in,obtain the high-order relations between nodes through patterns,and show a better ranking effect,but do not effectively design the characteristics of interaction frequency between nodes.With the rapid development of the information age,people have higher and higher requirements on the speed of information loading.A robust social network user ranking algorithm should not only return a list of users accurately for different purposes but also be fast and efficient.Based on the above problems,this paper based on the model of social network user ranking algorithm for research,based on considering the high-order relationship between user to introduce the concept of interaction frequency and the use of special sparse matrix storage method reduces the memory footprint,multi-core parallel computing CPU and GPU acceleration matrix iterative method shortened the experiment running time.I have completed the following work:(1)A motif-based interaction frequency weighted social network user ranking algorithm is proposed to deal with the problem of interaction frequency among users.This paper introduces the principle of weighting network graph and motif according to the frequency of interaction between users and discusses the ranking effect of three weighting methods to determine the optimal weighting method.To evaluate the performance of these methods,this paper conducts experiments on scholar networks and Twitter datasets and calculates the normalized discount cumulative gain of the motifbased PageRank algorithm and the three weighted methods with the ranking results of the approach centrality method as the idealized list.Their experimental effects under different coefficients were analyzed,and their results in the 3-node simple mode and the Anchor mode were compared.Experimental results show that the method of introducing motifs into the weighted social network graph has better accuracy than the motif-based PageRank algorithm.(2)A motif-based multi-process and GPU accelerated social network user ranking algorithm is proposed to reduce the algorithm memory and runtime.Firstly,the sparse rows and nested lists are compressed to reduce the memory occupied by the matrix.Then,the multi-process method is adopted to carry out continuous coding and matrix normalization parallel operation on the multi-core CPU.In the process of parallel matrix normalization,the non-zero elements in the sparse matrix are distributed in a top-heavy way.Furthermore,a dynamic slicing method based on the number of non-zero elements is proposed to make multiple processes end at approximately the same time,thus shortening the parallelization time.Finally,GPU is used to iterate the matrix.Compared with these methods,the proposed algorithm not only reduces the memory footprint but also shortens the running time.