Research On Abnormal User Detection Technology In Social Networks

During the prosperous development of the social network services,their users increasingly rely on the services owing to the openness of these services;however,a great many of anomalous users who go after the profits maliciously come into being.By spreading advertisement links and fake information,even making malicious attacks,they not merely degrade the services quality of the social network and disturb the social order,but also threaten the safety of common users severely and challenge the operations of the Public Security.The thesis comprehensively analyzes the academic fruits at home and abroad and the current research means,deeply researches the detection techniques of those anomalous users in some large-scale social networks such as Twitter and Microblogs,maps the anomalous detections of the social network to the classification in the vector space,clarifies the research idea of ensemble classification,and puts forward the improvement measures from the feature extraction and algorithm selection,so as to promote the effect of the anomalous detections.Firstly,thesis,in terms of the feature extraction,improves the current feature extraction methods of the social networks according to the fact that the current research fails to explore the attention relationship of users and puts forth the feature extraction model,so as to extract the selfnode feature and neighboring feature of users based on the information theory and homogeneity theory comprehensively.As for the node feature extraction,thesis applies some Natural Language Processing Models and complicated network calculation methods including Word2 vec and LDA etc,undertakes the full-sided extraction of the dominate features and hidden features in the aspects of texts and networks,and makes full use of the semantic information and network topology information.Besides,it also obtains the neighboring features based on weighting in the extraction of the multidimensional node features,selects K neighboring features with the most influential,which form the fusion feature set.By experimenting on the real and hand-annotated Weibo dataset,it also utilizes the four traditional classifiers to deal with the fusion feature set including SVM and the random forest,and the recall ratio can be promoted 5% maximally with effectiveness.Secondly,in the aspect of the algorithm selection,thesis firstly uses the gradient-dependent XGBoost algorithm as the detection method,which can be applied in the detection of the anomalous users in social networks,the optimization of the parameter selection and the establishment of the classification models,according to the deficiency of the conventional algorithm method to deal with the multi-dimensional features for the anomalous users in the social networks,information consumption caused by the feature selection,and low recall ratio and operation efficiency of its application in the unbalanced datasets.In accordance with reextraction feature of the datasets of the three real social networks,thesis sets up the unbalanced database with various percentages by the undersampling technique and makes use of XGBoost to identify the accounts which send many kinds of spam advertisements.It proves by the experiment results that the method can attain the outstanding performance in the recall ratio,which is up to 90% in the two-category and multi-category of the anomalous users in the social networks and it effectively promotes the recall ratio and F Value in comparison with the traditional classification methods including the random forest,and maintains better robustness in the balanced and unbalanced databases.