Research And Implementation Of Social Network Malicious User Identification And Detection

As a popular social networking tool,Weibo has attracted the use and participation of many netizens,accelerated the dissemination of information,and enhanced the economic value of the Internet.At the same time,the behavior of some malicious users has also caused harm to people.In addition,it has had a bad impact on the virtual network environment and the real society.How to effectively identify these malicious users has been an important issue and hot research direction in the construction and development of social networks.This article takes Sina Weibo as an object to study the method of identifying malicious users on Weibo,and to identify malicious users from the perspective of user feature analysis and building a learning model.The main work includes: First of all,get and organize the feature data set.On the basis of manual data annotation,semi-supervised clustering algorithm is used for data annotation,and the original feature extraction and purpose are analyzed.SVM-RFE,random forest-based feature selection method and PCA dimensionality reduction three feature analysis methods are introduced,and experiments are conducted one by one to select the best feature set of each method.Secondly,two machine learning models are introduced.First,based on the realization of the previously popular machine learning model,the XGBoost algorithm with outstanding integrated learning performance is introduced,and its principles and feature selection strategies are introduced.Second,based on the traditional fully connected neural network and correlation analysis of Weibo user characteristics,an integrated neural network based on feature association(INN-FA)is proposed,and the Adam optimizer is introduced.Realize the construction and verification of two types of models.By comparing ordinary machine learning algorithms and integrated learning models,and cross-experimenting with all the original feature sets of Weibo and the feature sets selected by different feature selection algorithms,it is concluded that integrated learning methods are generally superior to ordinary machine learning algorithms,and SVM-RFE + XGBoost combination method can get the best performance of 96.54% F value.In comparison of feature selection methods,the universality of SVM-RFE is better than that of random forest-based feature selection and PCA methods.For the comparison of neural network models,using the Tensorflow deep learning framework to build the Adam + INN-FA model,comparing the fully connected deep neural network before and after feature selection and the INN-FA model using traditional gradient descent methods.Adam + INNFA model reach an F value of 95.49%,which verified its effectiveness in recognition performance.At the same time,the unique advantage of this model is that it can reduce the structural complexity and enhance the flexibility of feature input.Finally,in order to further verify the universal applicability of the proposed model,the Twitter public data set was selected to verify the model.The experimental results show that using all 30 original features,the performance of the Adam + INN-FA model is comparable to that of the fully connected deep neural network.The XGBoost algorithm has the best performance on all data sets,slightly better than the random forest model proposed by the original author.Further feature selection for this data set,the SVM-RFE + XGBoost combination method that retains 8 important features with an F value of 97.68% can still outperform other feature selection combination models and other models using the original data set,which proves that effectiveness of the combined model.