CN112163848B - Role division system oriented to stream network, working method and medium thereof - Google Patents

Abstract

The invention relates to a role division system and its working method and medium, including a data acquisition module, a directional weighted network acquisition module, an embedded module and a clustering module; the data acquisition module is used to acquire transfer data; the directional weighted network acquisition module is used for The transfer data is expressed as a directed weighted network; the embedding module is used to first extract two undirected subgraphs for each node, then use the GraphWave algorithm to structurally embed, and finally integrate the structural embedding and the node’s in-out flow difference to obtain node embedding ; The clustering module uses the improved self-organizing map neural network to cluster the node embeddings obtained in the previous step to obtain the role division of the nodes. The invention can quickly discover the role composition of an economic organization, and find out the roles that may be senior members in combination with experience.

Description

A stream network-oriented role division system and its working method and medium

technical field

The invention relates to a role division system and its working method and medium, in particular to a flow network-oriented role division system and its working method.

Background technique

In the network, each node has its own role. The same role has similar behaviors, functions, and roles. For example, in an enterprise network, roles can be managers, team leaders, and ordinary employees. Roles help to study the key nodes of an organization or system, analyze the hierarchical structure, and provide reference information for comparison between multiple networks.

The role division of bank card numbers for an economic organization is conducive to understanding its organizational structure, and is conducive to quickly finding senior members and core members, and because only bank data is required, it will not attract the attention of the investigated organization. This is useful when investigating suspected illegal economic organizations such as pyramid schemes.

Currently, there is no public method to classify bank card numbers into roles. However, since transfer data can be expressed as a flow network, role division can also be expressed as a node role division of a convective network. Flow network is a special directed weighted network, its edges represent the flow of energy, matter, currency, information, etc., and the weight of the edges represents the flow. The flow network we focus on is an unbalanced flow network, which means that the incoming traffic of a node is not necessarily equal to the outgoing traffic. This is common in the real world, such as a network of money flows within a business. Some methods have been proposed at home and abroad on similar issues. However, since these methods can only classify the predefined three or four types of typical roles, or ignore the direction and weight of the network, they do not work well on directed weighted transfer networks.

The methods for dealing with approximation problems are as follows:

Tang Shiqi and others proposed a role discovery algorithm, which is essentially a node embedding algorithm. It uses the degree centrality, eigenvectors, and eigenvalues of nodes to construct a feature matrix, and then uses non-negative matrix decomposition. The object of processing is general network.

Li Wanyu proposed a role discovery algorithm. Inspired by the concept of physical field and potential, the concept of topological potential was proposed in the network, and for the directed weighted network, the out-topological potential and in-topological potential were proposed. However, the out- The (in) topological potential is just the sum of the influence and is not sensitive to the network structure. The algorithm can only distinguish four typical roles or fuzzy roles between the four roles. Such a role is literally the same as the role studied in this paper, but the substantive difference is huge.

Zheng Kunlun proposed a role analysis algorithm, which looks for important nodes according to some attributes of nodes (degree centrality, betweenness, PageRank ranking), then deletes important nodes, and then deletes only affected nodes, and then in the remaining To continue this process in the network, the important nodes are generally found by continuously dismantling the network. This algorithm does not consider the weight of the edges and is not suitable for weighted networks.

Zhong Xiaoyu proposed a role division method, which is similar to Li Wanyu's method mentioned above. This method can only distinguish three typical roles.

Node embedding (representation learning) is an important step in role division, and various node embedding methods have been proposed abroad. RolX, struc2vec and GraphWave are representation learning methods based on structural roles, and points with similar structural roles get close vectors. RolX uses the ReFeX algorithm to extract the feature matrix, and then uses non-negative matrix decomposition to obtain the embedding vector. struc2vec reconstructs the network according to the structural similarity of the nodes, connects two similar points with strong edges, and uses DeepWalk to obtain the vector representation of the nodes on the new network. struc2vec only pays attention to the topology, ignoring the attributes of edges and vertices. GraphWave uses diffusion wavelets to process Laplacian matrices, and treats wavelet coefficients as probability distributions. GraphWave has a good performance on undirected networks and can consider edge weights. RolX, struc2ve and GraphWave can only handle undirected graphs. DEG is a representation learning method on directed graphs, LINE, node2vec can be used for both undirected and directed graphs, but they are based on intimacy rather than similarity. Graph2Gauss uses network structure and node parameters to embed nodes, but the network structure it says only refers to distance, so nodes with close distances are more likely to get similar embedding vectors, so Graph2Gauss has nothing to do with roles. At present, there is no method for computing the node embedding of the flow network (representing the node as a vector), and then clustering and dividing the node embedding in the vector space. In this way, the node role division of the flow network is realized.

Contents of the invention

Aiming at the deficiencies of the prior art, the present invention provides a flow network-oriented role division system;

The invention also provides a working method of the role division system;

The purpose of the present invention is to divide a group of bank card numbers into roles, point out several roles in total, and which bank card numbers correspond to each role. The role division of the present invention cannot identify the meaning of each role, such as a boss or a manager. It only knows that bank card numbers belonging to the same role have similar transfer behaviors, and card numbers of different roles have different transfer behaviors. The role division of bank card numbers refers to dividing all the input bank card numbers into multiple groups. The bank card numbers in each group have the same role, and the roles of bank card numbers in different groups are different. For example, the input contains 5 bank card numbers a and b The data sets of , c, d, and e are processed by the role division system to obtain two groups, one group is abc, and the other group is de, which means that these card numbers have two roles, abc is a role, and de is a role.

Explanation of terms:

1. Convergence subgraph, for a point a, the convergence subgraph of point a is an undirected graph, the points of this graph include all the points of the original graph, and the edges of this graph include edges that may carry material flowing to point a. Use mathematical language to describe the generation process of the convergent subgraph edge set E _gat (a) of point a: For each edge <n, a> pointing to a in the original graph, add it to E _gat (a), and call n is the first-order upstream neighbor of a, do the same for each first-order upstream neighbor n of a, that is, add each edge <m, n> pointing to n to E _gat (a), and call m a’s The second-order upstream neighbors, and so on, add all the edges from the k+1-order upstream neighbors of a to the k-order upstream neighbors to E _gat (a), until no new edges are added, and the edges in the subgraph are aggregated The weight of the original image is retained, and the direction is discarded.

2. Diffusion subgraph, for a point a, the diffusion subgraph of point a is an undirected graph, the points of this graph include all the points of the original graph, and the edges of this graph include edges that may carry the material flowing out from point a. Use mathematical language to describe the generation process of the diffusion subgraph edge set E _dif (a) of a point: it is symmetric to the generation process of E _gat (a), and all the edges from the k-order downstream neighbors of a point to the k+1-order downstream neighbors Added to E _dif (a), until no new edges are added, the edges in the diffusion subgraph retain the weights of the original graph, and discard the direction.

3. PCA, Principal Components Analysis Principal component analysis, a well-known dimensionality reduction method.

Technical scheme of the present invention is:

A stream network-oriented role division system, including sequentially connected data acquisition module, directed weighted network acquisition module, embedding module and clustering module;

The data collection module is used to: obtain transfer data;

The directional weighted network acquisition module is used to: represent the transfer data as a directional weighted network;

The embedding module is used to: firstly extract two undirected subgraphs for each node, then use the GraphWave algorithm to structurally embed, and finally integrate the structural embedding and the difference between the incoming and outgoing traffic of the node to obtain the node embedding;

The clustering module is used to cluster the node embeddings obtained in the previous step by using the improved self-organizing map neural network to obtain the role division of the nodes.

The working method of the above flow network-oriented role division system includes the following steps:

(1) The data acquisition module obtains the transfer data; the transfer data refers to the transfer data between bank cards, and the transfer data between each bank card includes the card number of the transfer party, the card number of the transfer party, the amount, time;

(2) The directed weighted network acquisition module represents the transfer data as a directed weighted network;

(3) Obtain node embedding by the embedding module; first extract two undirected subgraphs for each node, then obtain structural embedding with GraphWave algorithm, and finally integrate structural embedding and node's flow difference to obtain node embedding;

(4) Realize role division through the clustering module: use the improved self-organizing map neural network to cluster the node embeddings obtained in the previous step to obtain the role division of the nodes.

Preferably according to the present invention, in step (2), expressing the transfer data as a directed weighted network means: expressing the card numbers of all transfer-out parties and all transfer-in parties as points in the directed weighted network, and transferring The cumulative transfer amount between the card number of the sender and the card number of the transferee is expressed as a directed edge between two points, from the transferer to the transferer, and the weight is the amount, that is, a directed weighted network.

Preferably according to the present invention, in step (3), extract two undirected subgraphs for each node, the undirected subgraphs include convergent subgraphs and diffusion subgraphs, including steps as follows:

G=(V, E), represents the original graph, that is, the directed weighted network obtained in step (2); V is a point set, including each point in the directed weighted network; E is an edge set, including the directed weighted network each edge in

The process of obtaining the converged subgraph is as follows:

G _gat (a)=(V, E _gat (a)), represents the converged subgraph about point a; point a is any point in the point set V; E _gat (a) refers to the edge of G _gat (a) set;

The calculation process of E _gat (a) is: for each edge <n, a> pointing to a, add it to E _gat (a), and call n the first-order upstream neighbor of a, for each a The first-order upstream neighbor n of a does the same treatment, that is, every edge <m, n> pointing to n is added to E _gat (a), and m is called the second-order upstream neighbor of a, and so on. The k+1-order upstream neighbor of a points to the k-order upstream neighbor's edge and adds it to E _gat (a) until no new edge is added, and the edge in the converged subgraph retains the weight of the original graph and discards the direction;

The acquisition process of the diffusion subgraph is as follows:

G _dif (a)=(V, E _dif (a)), represents the diffusion subgraph about point a; E _dif (a) refers to the edge set of G _dif (a);

The calculation process of E _dif (a) is: for each edge <a, n> whose source point is a, add it to E _dif (a), and call n the first-order downstream neighbor of a, for each Do the same for a first-order downstream neighbor n of a, that is, add each edge <n, m> with source point n to E _dif (a), and call m the second-order downstream neighbor of a, and so on , add all edges pointing to k+1-order upstream neighbors from the k-order downstream neighbors of a to E _dif (a), until no new edges are added, the edges in the diffusion subgraph retain the weight of the original graph, and the direction is discarded .

Preferably according to the present invention, in step (3), adopt GraphWave algorithm to obtain structural embedding, structural embedding is a vector, represents the position of point in the continuous space of structural character, comprises steps as follows:

The GraphWave algorithm is used to process G _gat (a), and the structural embedding _X a of point a in the convergence subgraph is obtained; the GraphWave algorithm is used to process G _dif (a), and the structural embedding Y _a of point a in the diffusion subgraph is obtained; Then the complete structural embedding of point a (X _a , Y _a ).

Preferably according to the present invention, in step (3), the node embedding is obtained by integrating the structural embedding and the difference between the incoming and outgoing traffic of the node, including the following steps:

The difference between incoming and outgoing traffic of a node is important information of a node in a flow network. Note that the weight of the side <i, j> is w _ij , i, j respectively represent the card number of the transfer-out party and the card number of the transfer-in party, and w _ij is the cumulative transfer amount from i to j; then, the inflow of point a f _in ＝∑w _ia , the outflow is f _out ＝∑w _ai , and the inflow and outflow difference is d＝f _in -f _out ;

In order to obtain a complete vector representation of a node, it is necessary to integrate the flow difference and structural embedding. However, the structural embedding is generally a vector of more than 16 dimensions (depending on the parameters). If it is directly put together with the difference in flow, the impact of the difference in flow will be very weak. It is necessary to reduce the structural embedding to 2 dimensions first, and then splicing it with the flow difference between in and out.

Reduce structural embeddings to 2 dimensions;

Further preferably, the structural embedding is reduced to 2 dimensions by PCA;

It is worth mentioning that, since the flow difference is actually used as the weight of the node, replace the incoming and outgoing flow difference with any other value that can be used as the weight of the node;

The structural embedding and the flow difference between in and out are spliced together to obtain a 3-dimensional vector representation. So far, for each point, a 3-dimensional vector representation is obtained, that is, node embedding.

Preferably according to the present invention, in step (4), use improved self-organizing map neural network SOM to carry out clustering to the node embedding obtained in previous step, obtain the role division of node, comprise steps as follows:

The improved self-organizing map neural network SOM includes an input layer and a competition layer, the competition layer and the input layer are fully connected, and the number of neurons in the input layer is the same as the dimension of a single input; the input layer neurons accept Input, competing layer neurons compete with each other.

The weight [w ₁ , w ₂ , w ₃ ......] of the improved self-organizing map neural network SOM is the position of the neuron in the competition layer; if the input layer has two neurons, then the neuron in the competition layer arranged on a two-dimensional plane.

A. Initialize dense competitive layer neurons;

B. Every time a stimulus p is input, the neural network calls the input a stimulus. The stimulus p refers to the node embedding obtained in step (3), and the shortest distance is selected from the competition layer, that is ||w _i -p|| _{2 is} the smallest The neuron i of is the winning neuron;

C. Adjust the winning neuron and its neighborhood N(i). Neighborhood N(i) refers to the position of each neuron in the area near i, so that they are close to p. Use α to represent the learning rate. Adjust the position as Shown in formula (I):

w _j ′=w _j +α(pw _j ), j∈N(i) (I)

In formula (I), j is a neuron in the neighborhood, w _j is the position of the neuron, and w _j ' is the adjusted new position;

After multiple trainings, neurons move into clusters;

D. Input two stimuli of the same cluster, corresponding to the same winning neuron, and the cluster center is the winning neuron;

A prominent advantage of SOM is that when it is used as a clustering algorithm, there is no need to specify the number of clusters, which is exactly the nature we need. However, multiple cluster centers often appear in a natural cluster. In order to solve this problem, the concept of resolution is proposed, which means that the weight of the competition layer can only take discrete values. In a discrete vector space, when a neuron moves to several inputs that are close together, it may move to the same position, which reduces the cluster center and improves the integrity. So how to choose the right resolution? Clearly, the optimal resolution varies from dataset to dataset, which introduces new problems.

E. Check all cluster centers, if two cluster centers are neighbors, or the neighbors of one cluster center and another cluster center overlap, then merge the two cluster centers, and the resulting cluster center is the role result of division.

Further preferably, the initialization of dense competitive layer neurons refers to: 4 neurons per unit area. That is, neurons are evenly distributed on a two-dimensional plane like the intersection points of a square grid. The side length of the square is 0.5, so there are 4 neurons per unit area.

Preferably according to the present invention, in step C, a higher resolution of 0.2 is used for training. It is equivalent to drawing a square grid on a two-dimensional plane. The side length of the square is 0.2. After the neuron adjusts its position each time, it needs to move to the nearest grid intersection. For example, the position is (0.12, 1.27), then move to (0.2, 1.2).

A computer-readable storage medium, characterized in that, the computer-readable storage medium stores a program of the working method of the stream network-oriented role division system, and the program of the working method of the stream network-oriented role division system is When executed by the processor, the steps of any one of the working methods of the flow network-oriented role division system are realized.

The beneficial effects of the present invention are:

1. The role division method proposed by the present invention can quickly discover the role composition of an economic organization, and find out the roles that may be senior members in combination with experience. If the investigated organization is an illegal organization, it can control the key bank accounts in the organization earlier investigation, improve the efficiency of case handling, reduce the loss of public property, and maintain economic order.

2. The present invention divides the roles of multiple economic organizations, can compare the role composition of these organizations, and is beneficial to identify special organizations whose role composition is different from ordinary organizations.

3. The practical scope includes all stream networks, and the application prospects include the role division of members of illegal economic organizations, the role division of transportation hubs, the identification of abnormal organizations based on role composition, and so on. Flow network is a special directed weighted network, its edges represent the flow of energy, matter, currency, information, etc., and the weight of the edges represents the flow.

Description of drawings

Figure 1 is a structural block diagram of a flow network-oriented role division system;

Fig. 2 is a schematic diagram of a working method of a role division system oriented to a flow network;

Figure 3(a) is a schematic diagram of the original picture;

Figure 3(b) is a schematic diagram of the converged subgraph for extracting point A from Figure (a);

Figure 3(c) is a schematic diagram of the diffusion subgraph for extracting point A from Figure (a).

Detailed ways

The present invention will be further limited below in conjunction with the accompanying drawings and embodiments, but not limited thereto.

Example 1

A flow network-oriented role division system, as shown in Figure 1, includes sequentially connected data acquisition modules, directed weighted network acquisition modules, embedding modules, and clustering modules; the data acquisition module is used to: acquire transfer data; directed The weighted network acquisition module is used to: represent the transfer data as a directed weighted network; the embedding module is used to: first extract two undirected subgraphs for each node, then use the GraphWave algorithm for structural embedding, and finally integrate the structural embedding and node The node embedding is obtained by the in-out traffic difference; the clustering module is used to: use the improved self-organizing map neural network to cluster the node embedding obtained in the previous step, and obtain the role division of the nodes.

Example 2

The working method of the flow network-oriented role division system described in Embodiment 1, as shown in Figure 2, includes the following steps:

(1) The data acquisition module obtains the transfer data; the transfer data refers to the transfer data between bank cards, and the transfer data between each bank card includes the card number of the transfer-out party, the card number of the transfer-in party, amount, and time;

(2) The directed weighted network acquisition module represents the transfer data as a directed weighted network; it means: express all the card numbers of the transfer-out party and all the card numbers of the transfer-in party as points in the directed weighted network, and the transfer-out party’s The cumulative transfer amount between the card number and the card number of the transferee is expressed as a directed edge between two points, from the transferee to the transferee, and the weight is the amount, that is, a directed weighted network.

(3) Obtain node embedding through the embedding module; first extract two undirected subgraphs for each node, then use GraphWave algorithm to obtain structural embedding, and finally integrate structural embedding and node in-out traffic difference to obtain node embedding;

(4) Realize role division through the clustering module: use the improved self-organizing map neural network to cluster the node embeddings obtained in the previous step, and obtain the role division of the nodes.

In step (3), two undirected subgraphs are extracted for each node, and the undirected subgraph includes a convergent subgraph and a diffusion subgraph, and the steps are as follows:

G=(V, E), represents the original image, as shown in Figure 3(a), that is, the directed weighted network obtained in step (2); V is a point set, including each point in the directed weighted network, including A, B, C, D, E, F, G, H, I, J; E is the edge set, including each edge in the directed weighted network;

As shown in Figure 3(b), the process of extracting the converged subgraph of point A is as follows:

G _gat (a)=(V, E _gat (a)), represents the converged subgraph about point a; point a is any point in the point set V; E _gat (a) refers to the edge of G _gat (a) set;

The calculation process of E _gat (a) is: for each edge <n, a> pointing to a, add it to E _gat (a), and call n the first-order upstream neighbor of a, for each a The first-order upstream neighbor n of a does the same treatment, that is, every edge <m, n> pointing to n is added to E _gat (a), and m is called the second-order upstream neighbor of a, and so on. The k+1-order upstream neighbor of a points to the k-order upstream neighbor's edge and adds it to E _gat (a) until no new edge is added, and the edge in the converged subgraph retains the weight of the original graph and discards the direction;

As shown in Figure 3(c), the acquisition process of extracting the diffusion subgraph of point A is as follows:

G _dif (a)=(V, E _dif (a)), represents the diffusion subgraph about point a; E _dif (a) refers to the edge set of G _dif (a);

The calculation process of E _dif (a) is: for each edge <a, n> whose source point is a, add it to E _dif (a), and call n the first-order downstream neighbor of a, for each Do the same for a first-order downstream neighbor n of a, that is, add each edge <n, m> with source point n to E _dif (a), and call m the second-order downstream neighbor of a, and so on , add all edges pointing to k+1-order upstream neighbors from the k-order downstream neighbors of a to E _dif (a), until no new edges are added, the edges in the diffusion subgraph retain the weight of the original graph, and the direction is discarded .

In step (3), the GraphWave algorithm is used to obtain the structural embedding. The structural embedding is a vector representing the position of the point in the continuous space of the structural role, including the following steps:

The GraphWave algorithm is used to process G _gat (a), and the structural embedding _X a of point a in the convergence subgraph is obtained; the GraphWave algorithm is used to process G _aif (a), and the structural embedding Y _a of point a in the diffusion subgraph is obtained; Then the complete structural embedding of point a (X _a , Y _a ).

In step (3), the node embedding is obtained by integrating the structural embedding and the difference between the incoming and outgoing flow of the node, including the following steps:

The difference between incoming and outgoing traffic of a node is important information of a node in a flow network. Note that the weight of the side <i, j> is w _ij , i, j respectively represent the card number of the transfer-out party and the card number of the transfer-in party, and w _ij is the cumulative transfer amount from i to j; then, the inflow of point a f _in ＝∑w _ia , the outflow is f _out ＝∑w _ai , and the inflow and outflow difference is d＝f _in -f _out ;

In order to obtain a complete vector representation of a node, it is necessary to integrate the flow difference and structural embedding. However, the structural embedding is generally a vector of more than 16 dimensions (depending on the parameters). If it is directly put together with the difference in flow, the impact of the difference in flow will be very weak. It is necessary to reduce the structural embedding to 2 dimensions first, and then splicing it with the flow difference between in and out.

Reduce the structural embedding to 2 dimensions by PCA;

It is worth mentioning that, since the flow difference is actually used as the weight of the node, replace the incoming and outgoing flow difference with any other value that can be used as the weight of the node;

The structural embedding and the flow difference between in and out are spliced together to obtain a 3-dimensional vector representation. So far, for each point, a 3-dimensional vector representation is obtained, that is, node embedding.

In step (4), use the improved self-organizing map neural network SOM to cluster the node embeddings obtained in the previous step, and obtain the role division of the nodes, including the following steps:

The improved self-organizing map neural network SOM includes an input layer and a competition layer, and the competition layer and the input layer are fully connected, and the number of neurons in the input layer is the same as the dimension of a single input; the input layer neurons accept input, Competitive layer neurons compete with each other.

The weight [w ₁ , w ₂ , w ₃ ......] of the improved self-organizing map neural network SOM is the position of the neuron in the competition layer; if the input layer has two neurons, then the neuron in the competition layer arranged on a two-dimensional plane.

A. Initialize dense competitive layer neurons;

B. Every time a stimulus p is input, the neural network calls the input a stimulus. The stimulus p refers to the node embedding obtained in step (3), and the shortest distance is selected from the competition layer, that is ||w _i -p|| _{2 is} the smallest The neuron i of is the winning neuron;

C. Adjust the winning neuron and its neighborhood N(i). Neighborhood N(i) refers to the position of each neuron in the area near i, so that they are close to p. Use α to represent the learning rate. Adjust the position as Shown in formula (I):

w _j ′=w _j +α(pw _j ), j∈N(i) (I)

In formula (I), j is a neuron in the neighborhood, w _j is the position of the neuron, and w _j ' is the adjusted new position;

After multiple trainings, neurons move into clusters;

D. Input two stimuli of the same cluster, corresponding to the same winning neuron, and the cluster center is the winning neuron;

A prominent advantage of SOM is that when it is used as a clustering algorithm, there is no need to specify the number of clusters, which is exactly the nature we need. However, multiple cluster centers often appear in a natural cluster. In order to solve this problem, the concept of resolution is proposed, which means that the weight of the competition layer can only take discrete values. In a discrete vector space, when a neuron moves to several inputs that are close together, it may move to the same position, which reduces the cluster center and improves the integrity. So how to choose the right resolution? Clearly, the optimal resolution varies from dataset to dataset, which introduces new problems.

E. Check all cluster centers, if two cluster centers are neighbors, or the neighbors of one cluster center and another cluster center overlap, then merge the two cluster centers, and the resulting cluster center is the role result of division.

Initialize dense competitive layer neurons, which means: 4 neurons per unit area. That is, neurons are evenly distributed on a two-dimensional plane like the intersection points of a square grid. The side length of the square is 0.5, so there are 4 neurons per unit area.

In step C, take a higher resolution of 0.2 for training. It is equivalent to drawing a square grid on a two-dimensional plane. The side length of the square is 0.2. After the neuron adjusts its position each time, it needs to move to the nearest grid intersection. For example, the position is (0.12, 1.27), then move to (0.2, 1.2).

Example 3

A computer-readable storage medium, the computer-readable storage medium stores the program of the working method of the flow network-oriented role division system described in Embodiment 2, and the working method of the flow network-oriented role division system described in Embodiment 2 When the program is executed by the processor, the steps of any working method of the stream network-oriented role division system described in Embodiment 2 are realized.

Claims (8)

1. A method of work of a stream network-oriented role division system, characterized in that, comprising successively connected data acquisition modules, directed weighted network acquisition modules, embedding modules and clustering modules; The data acquisition module is used to: obtain transfer data; the directed weighted network acquisition module is used to: represent the transfer data as a directed weighted network; the embedded module is used to: first extract two undirected subgraph, and then use the Graphwave algorithm to structurally embed, and finally integrate the structural embedding and the difference between the incoming and outgoing traffic of the node to obtain the node embedding; the clustering module is used to: use the improved self-organizing map neural network to carry out the node embedding obtained in the previous step Clustering to obtain the role division of nodes; the steps are as follows: (1) The data acquisition module obtains the transfer data; the transfer data refers to the transfer data between bank cards, and the transfer data between each bank card includes the card number of the transfer party, the card number of the transfer party, the amount, time; (2) The directed weighted network acquisition module represents the transfer data as a directed weighted network; (3) Obtain node embedding by the embedding module; first extract two undirected subgraphs for each node, then obtain structural embedding with GraphWave algorithm, and finally integrate structural embedding and node's flow difference to obtain node embedding; (4) Realize the role division by the clustering module: use the improved self-organizing map neural network to cluster the node embedding obtained in the previous step, and obtain the role division of the node; In step (3), two undirected subgraphs are extracted for each node. The undirected subgraph includes a convergent subgraph and a diffusion subgraph, and the steps are as follows: G=(V, E), represents the original graph, that is, the directed weighted network obtained in step (2); V is a point set, including each point in the directed weighted network; E is an edge set, including the directed weighted network each edge in The process of obtaining the converged subgraph is as follows: G _gat (a)=(V, E _gat (a)), represents the converged subgraph about point a; point a is any point in the point set V; E _gat (a) refers to the edge of G _gat (a) set; The calculation process of E _gat (a) is: for each edge <n, a> pointing to a, add it to E _gat (a), and call n the first-order upstream neighbor of a, for each a The first-order upstream neighbor n of a does the same treatment, that is, every edge <m, n> pointing to n is added to E _gat (a), and m is called the second-order upstream neighbor of a, and so on, all from The k+1-order upstream neighbor of a points to the k-order upstream neighbor's edge and adds it to E _gat (a) until no new edge is added, and the edge in the converged subgraph retains the weight of the original graph and discards the direction; The acquisition process of the diffusion subgraph is as follows: G _dif (a)=(V, E _dif (a)), represents the diffusion subgraph about point a; E _dif (a) refers to the edge set of G _dif (a); The calculation process of E _dif (a) is: for each edge <a, n> whose source point is a, add it to E _dif (a), and call n the first-order downstream neighbor of a, for each A first-order downstream neighbor n of a does the same process, that is, add each edge <n, m> with source point n to E _dif (a), and call m the second-order downstream neighbor of a, and so on , add all edges pointing to k+1-order upstream neighbors from the k-order downstream neighbors of a to E _dif (a), until no new edges are added, the edges in the diffusion subgraph retain the weight of the original graph, and the direction is discarded . 2. The working method of the stream network-oriented role division system according to claim 1, characterized in that, in step (2), expressing the transfer data as a directed weighted network means: all the card numbers of the transferring party and all transfer-in party's card numbers are expressed as points in the directed weighted network, the cumulative transfer amount between the transfer-out party's card number and the transfer-in party's card number is expressed as a directed edge between two points, from the transfer-out party Point to the transferee, and the weight is the amount, that is, a directed weighted network. 3. The working method of the stream network-oriented role division system according to claim 1, characterized in that, in step (3), the GraphWave algorithm is used to obtain the structural embedding, and the structural embedding is a vector representing points in the structural The position of the continuous space of the character, including the following steps: The GraphWave algorithm is used to process G _gat (a), and the structural embedding _X a of point a in the convergence subgraph is obtained; the GraphWave algorithm is used to process G _dif (a), and the structural embedding Y _a of point a in the diffusion subgraph is obtained; Then the complete structural embedding of point a (X _a , Y _a ). 4. the working method of the role division system oriented flow network according to claim 1, is characterized in that, in step (3), the flow difference of integration structural embedding and node obtains node embedding, comprises steps as follows: Note that the weight of the edge <i, j> is w _ij , i, j represent the card number of the transfer-out party and the card number of the transfer-in party, respectively, and w _ij is the cumulative transfer amount from i to j; then, the inflow of point a f _in ＝∑w _ia , the outflow is f _out ＝∑w _ai , and the inflow and outflow difference is d＝f _in -f _out ; Reduce structural embeddings to 2 dimensions; The structural embedding and the flow difference between in and out are spliced together to obtain a 3-dimensional vector representation. So far, for each point, a 3-dimensional vector representation is obtained, that is, node embedding. 5. The working method of the flow network-oriented role division system according to claim 4, characterized in that, the structural embedding is reduced to 2 dimensions by PCA. 6. according to the working method of the described flow network-oriented role division system of claim 1-5, it is characterized in that, in step (4), use improved self-organizing map neural network SOM to the node embedding that last step obtains Perform clustering to obtain the role division of nodes, including the following steps: The improved self-organizing map neural network SOM includes an input layer and a competition layer, the competition layer and the input layer are fully connected, and the number of neurons in the input layer is the same as the dimension of a single input; The weight [w ₁ , w ₂ , w ₃ ......] of the improved self-organizing map neural network SOM is the position of the neuron in the competition layer; A. Initialize dense competitive layer neurons; B. Each time a stimulus p is input, the stimulus p refers to the node embedding obtained in step (3), and the neuron i with the shortest distance, i.e. ||w _i -p|| ₂ , is selected from the competition layer as the winning neuron Yuan; C. Adjust the winning neuron and its neighborhood N(i). Neighborhood N(i) refers to the position of each neuron in the area near i, so that they are close to p. Use α to represent the learning rate. Adjust the position as Shown in formula (I): w _j ′=w _j +α(pw _j ), j∈N(i) (I) In formula (I), j is a neuron in the neighborhood, w _j is the position of the neuron, and w _j ' is the adjusted new position; After multiple trainings, neurons move into clusters; D. Input two stimuli of the same cluster, corresponding to the same winning neuron, and the cluster center is the winning neuron; E. Check all cluster centers, if two cluster centers are neighbors, or the neighbors of one cluster center and another cluster center overlap, then merge the two cluster centers, and the resulting cluster center is the role result of division. 7. The working method of the flow network-oriented role division system according to claim 6, characterized in that, the initialization of dense competitive layer neurons refers to: 4 neurons per unit area. 8. A computer-readable storage medium, characterized in that, the computer-readable storage medium stores the program of the working method of any stream network-oriented role division system according to claim 1-7, and the stream network-oriented When the program of the working method of the role division system is executed by the processor, the steps of the working method of the stream network-oriented role division system described in any one of claims 1-7 are realized.

Images