Detecting Abnormal Accounts On Large-scale Social Networks Based On Pregel Graph Computation

The detection of abnormal accounts in online social networks is one of the key issues that need to be improved in the development of digital economy.The abnormal accounts often severely affect the user experience of the online social network and the social credit system through bad Internet behavior.At present,the detection of abnormal account in online social network can be mainly divided into three schemes,supervised detection,semi-supervised detection and unsupervised detection.The supervised detection needs to train classifiers in advance,and it is difficult to detect unknown behaviors.Semi-supervised detection cannot correct its own errors,resulting in very unstable detection results.Unsupervised detection schemes based on graph structure may be a more promising way to detect abnormal accounts,but the low accuracy rate and long calculation time-consuming lose timeliness are bottlenecks that limits its large-scale application.Therefore,an abnormal account detection algorithm based on the node importance had been proposed in online social network in this paper.By including the node importance directly related to the network structure into the power iteration process in the SybilRank algorithm,the attributes of the node itself are retained to a greater extent.Furthermore,according to the current research status shows that there is no method to use association relationship in the detection of abnormal accounts up to now.This paper proposes for the first time that abnormal accounts in online social networks can be detected through the relativity of friends.Besides,a weighted compression formula for establishing the importance coefficient of nodes is proposed.This paper improves the original distributed parallel system of the algorithm.We select to use Pregel system to replace the original Map Reduce system,which is a pointcentric distributed parallel computing system based on the BSP model,it is more suitable for graph parallel computing and helpful for huge implementation of abnormal account detection in online social networks.We tested multiple datasets of social network at different scales,after obtaining the best parameters according to the Receiver Operating Characteristic(ROC)curve,we evaluated the detection effect through the Area Under Curve(AUC)and Matthews correlation coefficient(MCC)from the whole point and the extreme point.We found that the accuracy of our improved algorithm has been greatly improved.Moreover,with the increase of the proportion of attack edges,the failure rate of our improved algorithms become slower.It is worth mentioning that when the proportion of attack edges is relatively small,the value of MCC obtained by our improved algorithms are almost close to 1.By the way,the distributed parallel graph computing framework Pregel can reduce the computing time-consuming cost by around thirty times compared with the original framework.In addition,we compare the effect of single-node parallelism with multi-node distributed parallelism in order to find computing peaks to ensure maximum computing performance and utilization of computing resources.