Social Network Forensics Analysis Model Based On Network Representation Learning

With the continuous development of computer technology and social network,people’s communication left a lot of data on the network,including a large number of information,information mining and relationship research has become an important topic.The purpose of social network forensics is to preserve,analyze and provide objective,fair,direct and effective third-party evidence from people’s communication data.By analyzing the data of criminal mobile communication,we can analyze the relationship between members and the whole criminal network,and get the core members and even the heads of the crime.Through the tracking and control of these members,we can fight against the whole criminal network and get good results.The traditional forensic analysis can be based on Web forensic analysis and community structure algorithm.This paper briefly introduces the construction and process of the basic social network forensics model,and introduces the basic concepts of social network research,including: the centrality to measure the importance of nodes in the network and similarity calculation method and so on.Secondly,I describe the basic characteristics of social networks for later analysis and model building,which are: small world characteristic,scale-free characteristic,community structure characteristic and hierarchical characteristic;finally,I describe the basis of network representation learning algorithm and three classical representation learning algorithms,deepwalk,line and node2 vec,and make comparative analysis.Mapping criminal network to vector space with node2 vec algorithm through comparative analysis.The information of node attributes and network structure is reserved reasonably,which is convenient to calculate the relationship between nodes.At the same time,the process of random walk is improved,and BFS and DFS are used to control the walk mode of nodes.In this paper,a forensics model SNFA based on network representation learning is proposed.First of all,based on the characteristics of social networks,I make the selection rules of computable nodes in the network.Then,the sampling and coding process of network nodes is introduced,and the vectorization representation of nodes is realized by using CBOW model and hierarchical softmax model.At the same time,according to the characteristics of social networks,the random walk mode of node sampling is improved,which tends to walk to the same level of intimate nodes.The gradient updating process of CBOW model based on the average random gradient rising method is given.Finally,I briefly describe the clustering method.In the criminal network,first use the hierarchical clustering to get the best cluster center value in the whole clustering process,then carry out iterative relocation to achieve the best clustering,and get the core nodes set of social network.By using cosine distance formula and Euclidean distance formula,the importance degree of each node in the core nodes set is calculated,and the core nodes sequence based on the importance degree is obtained by descending sorting.The main work of this paper is as follows:(1)Because the traditional community structure algorithm has some disadvantages,this paper uses the network representation learning method,maps the criminal network into vector space,reasonably expresses the node attributes and network structure in the network,uses the node2 vec algorithm to represent the network nodes as vector form,and improves the random walk mode.(2)In this paper,the random walk of sampling process is improved,and the transition probability is used to make the random walk tend to the same level and intimate nodes.At the same time,the gradient updating process of CBOW model based on the average random gradient rising method is given,so that each update value is evenly distributed to each node.(3)According to the characteristics of social networks,the hierarchical clustering method is used to obtain the complete clustering structure,the optimal cluster center value is obtained,and the iterative relocation method is used to obtain the optimal cluster.The importance degree of each node in the core node set is calculated by the method of combining two distance calculation angles,and the core node sequence based on the importance degree is obtained by descending sorting.