社交网络中协同舆论欺诈检测方法应用研究

Research on Application of Collaborative Public Opinion Fraud Detection Method in Social Network

为了能够使网络空间提供更加可靠的信息,欺诈检测变得越来越重要,但现有的方法在检测欺诈用户时仅考虑了用户评论之间评论相同商品时形成的静态密集子图,而忽略了用户自身在评论时的异常行为,从而导致准确性较低,在实践中往往需要进一步手动验证检测结果的可靠性。针对此问题,提出了一种协同舆论欺诈检测(CPOFD)方法,该方法使用一种新的度量,即对比可疑度。该度量主要包括拓扑连接的信息,使得CPOFD方法能够通过拓扑连接、时间戳以及评分等信息有效检测欺诈者的异常行为,以更为聚合的方式检测欺诈群体。该度量强调了欺诈者和正常用户的动态对比,使得算法能够在拓扑连接、时间戳以及评分方面更为有效地检测欺诈者的异常行为。同时,CPOFD方法结合基于密度子图的聚类算法和决策树分类算法将社交网络中用户进行有效分组,且在聚簇分类时使用模拟退火算法进行剪枝优化,能更加简洁快速地寻找近似最优解,时间复杂度与欺诈者数量呈线性关系,具有较高的可扩展性。基于Yelp数据集的实验结果表明:CPOFD方法对欺诈舆论检测的准确度大多数在98%以上,验证了CPOFD方法的有效性。

In order to ensure cyberspace to provide more reliable information,fraud detection became more important.Existing methods only considered the static dense sub-graphs formed between user comments when detecting fraudulent users,while ignored the abnormal behavior of users during the comments,which led to reduced accuracy.Meanwhile,further manual verification was often required to verify the reliability of the results in practice.For this problem,this paper proposed the CPOFD method,which used a new measure“comparative equivocation”.This measure mainly included topological connection information to detect fraud groups in a more aggregated manner.Specifically,this metric emphasized the dynamic comparison between fraudsters and normal users,so that the algorithm could effectively detect the fraudster′s abnormal behavior in terms of topological connections,timestamps,and scoring information.At the same time,this method combined the clustering algorithm based on the density sub-graphs and the decision tree classification algorithm to group users in the social network effectively,and used the simulated annealing algorithm to optimize the pruning when classifying the clusters,so as to find the approximate optimum solution more concisely and quickly.The time complexity of the algorithm was linear to the number of fraudsters,and it had high scalability.In experiments based on the Yelp dataset,the accuracy of the CPOFD method for fraudulent public opinion detection reached more than 98%,which verified the effectiveness of the CPOFD method.