CN114066081B - Enterprise risk prediction method and device based on graph attention network and electronic equipment - Google Patents

Enterprise risk prediction method and device based on graph attention network and electronic equipment Download PDF

Info

Publication number
CN114066081B
CN114066081B CN202111394584.1A CN202111394584A CN114066081B CN 114066081 B CN114066081 B CN 114066081B CN 202111394584 A CN202111394584 A CN 202111394584A CN 114066081 B CN114066081 B CN 114066081B
Authority
CN
China
Prior art keywords
network
enterprise
weight
attention
graph
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202111394584.1A
Other languages
Chinese (zh)
Other versions
CN114066081A (en
Inventor
李若愚
高佳月
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Hengtong Huiyuan Big Data Technology Co ltd
Original Assignee
Beijing Hengtong Huiyuan Big Data Technology Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Hengtong Huiyuan Big Data Technology Co ltd filed Critical Beijing Hengtong Huiyuan Big Data Technology Co ltd
Priority to CN202111394584.1A priority Critical patent/CN114066081B/en
Publication of CN114066081A publication Critical patent/CN114066081A/en
Application granted granted Critical
Publication of CN114066081B publication Critical patent/CN114066081B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0635—Risk analysis of enterprise or organisation activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/018—Certifying business or products

Landscapes

  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Engineering & Computer Science (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Development Economics (AREA)
  • General Physics & Mathematics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Game Theory and Decision Science (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Educational Administration (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an enterprise risk prediction method and device based on a graph attention network and electronic equipment. The method comprises the following steps: constructing a network diagram of the relationship of the enterprises with the weight and the association relationship of the enterprises with the weight; taking the network diagram of the association relationship of the authorized enterprises and the state information of each enterprise in the network diagram at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the network of attention by the diagram; in the graph attention network, attention weights of nodes are calculated by using a weighted enterprise incidence relation, wherein the weights of edges among the nodes are determined by expert fitting according to the incidence relation among the nodes and based on financial business knowledge. The risk states of the enterprises and the related enterprises are predicted by adopting the graph attention network, and the result is accurate; by adding incidence relation weight correction based on business knowledge in attention score calculation, model interpretability is enhanced, and accuracy of a prediction result is further improved.

Description

Enterprise risk prediction method and device based on graph attention network and electronic equipment
Technical Field
The invention relates to the technical field of enterprise risk prediction, in particular to an enterprise risk prediction method and device based on a graph attention network and electronic equipment.
Background
Enterprise risk prediction refers to assessing the likelihood of risk occurrence for other enterprises associated (directly or indirectly) with the enterprise at risk occurrence based on inter-enterprise relationships. The enterprise risk prediction is of great significance in practical application, for example, in a commercial bank, the prediction of the loan default state of an enterprise client can help the bank to reasonably loan and reduce the rate of bad loans. The investor can also adopt different investment strategies according to the predicted enterprise risk, and the loss caused by error investment is reduced. Therefore, the risk prediction of the enterprise can not only create more value for the financial institution, but also bring better return for the individual.
At present, machine learning methods such as logistic regression and the like and time sequence analysis methods such as long-short term memory networks, Markov networks and the like are generally adopted for enterprise risk prediction, but the enterprise risk analysis method ignores the incidence relation between enterprises and the network structure, loses a lot of important enterprise incidence information, causes inaccurate prediction results and has larger errors.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides the following technical scheme.
The invention provides an enterprise risk prediction method based on a graph attention network, which comprises the following steps:
constructing a network diagram of the relationship of the enterprises with the weight and the association relationship of the enterprises with the weight;
taking the network diagram of the association relationship of the authorized enterprises and the state information of each enterprise in the network diagram at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the network of attention by the diagram; wherein, in the graph attention network, the attention weight of the node is calculated by using the association relation of the enterprise with the weight;
in the network diagram of the incidence relation of the enterprises with the weight, the enterprises are used as nodes, and the incidence relation between the enterprises with the weight is used as an edge; the weight of the edge between the nodes i and j is determined by expert fitting according to the incidence relation between the nodes i and j and based on financial business knowledge;
the graph attention network sequentially comprises a linear layer, two graph attention layers and an output layer;
in the graph attention layer, the graph attention layer is thinned by using a mode of multiplying the weight of the edge and the attention weight, and the attention matrix of the node is calculated according to the following formula:
Figure BDA0003369537050000021
Figure BDA0003369537050000022
wherein the content of the first and second substances,
Figure BDA0003369537050000023
represents nodes i and j tol-layer attention weight;
Figure BDA0003369537050000024
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure BDA0003369537050000025
representing the ith layer feature vector of the node i;
Figure BDA0003369537050000026
representing the ith layer feature vector of the node j;
Figure BDA0003369537050000027
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weights of the edges of nodes i and j.
Further, the constructing a weighted enterprise association relationship network graph with weighted relationships among the enterprises comprises:
acquiring enterprise association relation data;
and taking the enterprises as nodes, and taking the incidence relation among the enterprises with the weights as edges to obtain a network graph of the incidence relation of the enterprises with the weights.
Further, the network diagram of the incidence relation of the enterprise with the right is represented by a adjacency matrix with the right, and the adjacency matrix with the right is used as the input of the attention network of the diagram, wherein the adjacency matrix with the right is as follows:
Figure BDA0003369537050000031
wherein A [ i ] [ j ] is a weighted adjacency matrix, i, j are nodes in the network, E is a set of edges in the network, w is a weight of an edge, and
Figure BDA0003369537050000032
wherein R isijIs a set of relationships between nodes i and j, rijIs the relation weight between the nodes i and j, and n is the relation number between the nodes i and j.
Further, the relationship between the nodes i and j comprises a debt type, an equity type, a supply type and a macro type, and the debt type relationship comprises a guarantee providing relationship, a guaranteed relationship and an inter-insurance joint guarantee; the stock right type relationship comprises a stock control relationship, a stock participation relationship, a relationship controlled by a third enterprise, a common actual controller, a relationship of relatives of the actual controller and a common high-level manager; the supply-type relationship comprises a supply chain; the macro-type relationship comprises the same industry and the same region.
Further, the state information of the enterprise includes: enterprise basic information, financial indexes and early warning factors.
In another aspect, the present invention provides an enterprise risk prediction device based on a graph attention network, including:
the system comprises a weighted enterprise incidence relation network graph construction module, a weighted enterprise incidence relation network graph generation module and a weighted enterprise incidence relation network graph generation module, wherein the weighted enterprise incidence relation network graph construction module is used for constructing a weighted enterprise incidence relation network graph with inter-enterprise relations; in the network diagram of the incidence relation of the enterprises with the weight, the enterprises are used as nodes, and the incidence relation between the enterprises with the weight is used as an edge; the weight of the edge between the nodes i and j is determined by expert fitting according to the incidence relation between the nodes i and j and based on financial business knowledge;
the graph attention network prediction module is used for taking the authorized enterprise incidence relation network graph and the state information of each enterprise in the network graph at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the graph attention network; wherein, in the graph attention network, the attention weight is calculated by using the association relation of the enterprise with the weight;
the graph attention network sequentially comprises a linear layer, two graph attention layers and an output layer;
in the graph attention layer, the graph attention layer is thinned by using a mode of multiplying the weight of the edge and the attention weight, and the attention matrix of the node is calculated according to the following formula:
Figure BDA0003369537050000041
Figure BDA0003369537050000042
wherein the content of the first and second substances,
Figure BDA0003369537050000043
representing the l-th layer attention weight of the nodes i and j;
Figure BDA0003369537050000044
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure BDA0003369537050000045
representing the ith layer feature vector of the node i;
Figure BDA0003369537050000046
representing the ith layer feature vector of the node j;
Figure BDA0003369537050000047
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weights of the edges of nodes i and j.
A third aspect of the invention provides a memory storing a plurality of instructions for implementing the method described above.
A fourth aspect of the present invention provides an electronic device, comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, and the instructions are loaded and executed by the processor, so that the processor can execute the method.
The invention has the beneficial effects that: the embodiment of the invention provides an enterprise risk prediction method and device based on a graph attention network and electronic equipment. Firstly, constructing a network diagram of the association relationship of enterprises with rights, wherein the relationships among the enterprises have weights; then taking the network diagram of the association relationship of the authorized enterprises and the state information of each enterprise in the network diagram at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the network of attention; wherein, in the graph attention network, the attention weight of the node is calculated by using the enterprise association relation with the weight. The method and the system aggregate the associated information by adopting the graph attention network, predict the risk states of the enterprises and the associated enterprises, and ensure accurate prediction results due to the consideration of the association relationship among the enterprises and the structure of the network; in addition, in the calculation of the attention score, the interpretability of the model is enhanced by adding the incidence relation weight correction based on business knowledge, and the accuracy of the prediction result is further improved.
Drawings
FIG. 1 is a schematic flow chart of an enterprise risk prediction method based on a graph attention network according to the present invention;
fig. 2 is a schematic diagram of an enterprise risk prediction device based on a graph attention network according to the present invention.
Detailed Description
In order to better understand the technical solution, the technical solution will be described in detail with reference to the drawings and the specific embodiments.
The method provided by the invention can be implemented in the following terminal environment, and the terminal can comprise one or more of the following components: a processor, a memory, and a display screen. Wherein the memory has stored therein at least one instruction that is loaded and executed by the processor to implement the methods described in the embodiments described below.
A processor may include one or more processing cores. The processor connects various parts within the overall terminal using various interfaces and lines, performs various functions of the terminal and processes data by executing or executing instructions, programs, code sets, or instruction sets stored in the memory, and calling data stored in the memory.
The Memory may include a Random Access Memory (RAM) or a Read-Only Memory (ROM). The memory may be used to store instructions, programs, code sets, or instructions.
The display screen is used for displaying user interfaces of all the application programs.
In addition, those skilled in the art will appreciate that the above-described terminal configurations are not intended to be limiting, and that the terminal may include more or fewer components, or some components may be combined, or a different arrangement of components. For example, the terminal further includes a radio frequency circuit, an input unit, a sensor, an audio circuit, a power supply, and other components, which are not described herein again.
Example one
As shown in fig. 1, an embodiment of the present invention provides an enterprise risk prediction method based on a graph attention network, including:
s101, constructing a weighted enterprise incidence relation network graph with the inter-enterprise relation having weight;
s102, taking the network diagram of the association relationship of the authorized enterprises and the state information of each enterprise in the network diagram at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the network of attention by the diagram; wherein, in the graph attention network, the attention weight of the node is calculated by using the enterprise association relation with the weight.
In step S101, the enterprise association relationship network graph is composed of nodes and edges, and represents association relationships between node objects and each other. Today, many important data exist in the form of network graphs, such as a fund relationship network formed among enterprises, a citation network formed among academic papers, a social network formed among social users, and the like. For different types of networks, nodes and edges may represent different meanings. Meanwhile, in order to enrich the information of the network graph, the nodes and the edges can also have various attributes, for example, in an enterprise association relationship network graph, the nodes represent enterprises, and the edges represent relationships among the enterprises, such as investment, guarantee, transaction and the like. Each enterprise node may also contain financial information, industry information, etc. of the enterprise itself. For a network graph containing node state information, the node state generally changes with time and is time-sequential. To better represent the inter-enterprise relationships, intensity correction may also be added on the edge, i.e., the weight of the inter-enterprise relationships is added on the edge. The network graph with the relationship weight between the enterprises increased on the edge is the associated relationship network graph of the enterprises with the weight.
In the actual implementation process, step S101 may be implemented as follows:
acquiring enterprise association relation data;
and taking the enterprises as nodes, and taking the incidence relation among the enterprises with the weights as edges to obtain an enterprise incidence relation network graph with the weights.
The enterprise association relation data set can be derived from public enterprise association relation data and transaction data in banks, and the data can include enterprise financial statement information, risk rating (in-line and out-of-line), investment, guarantee and transaction data and the like of each year. After the data is processed, the enterprises can be used as nodes, and the incidence relation among the enterprises with the weights is used as an edge, so that an enterprise incidence relation network graph with the weights is obtained. The edges with weights can be obtained through expert fitting according to the incidence relation among enterprises.
In the embodiment of the invention, the relations between enterprises can be divided into four types, including debt type, stock right type, supply type and macro type, wherein the debt type relations include offering guarantee relations, guaranteed relations and inter-insurance joint guarantee bodies; the stock right type relationship comprises a stock control relationship, a stock participation relationship, a relationship controlled by a third enterprise, a common actual controller, a relationship of relatives of the actual controller and a common high-level manager; the supply-type relationship comprises a supply chain; the macro-type relationship comprises the same industry and the same region.
In a preferred embodiment of the present invention, the relationship weight between enterprises is determined based on the relationship between enterprises.
The inter-enterprise relationship types, inter-enterprise relationships, and relationship weights may be found in the following table.
Figure BDA0003369537050000071
Figure BDA0003369537050000081
In the above table, the inter-enterprise relationship weight can be obtained by expert fitting as follows:
by analyzing the incidence relation among enterprises in which the risk conduction occurs, the data shows that the enterprise risk conduction is mainly influenced by the incidence relation of the debt type, the stock right type, the supply type and the macro type. And taking the relation as a characteristic, dividing the enterprise samples into two categories of occurrence conduction and normal, and solving a characteristic weight coefficient by using an XGboost ensemble learning model. Since the overfitting condition can be caused by carrying out weight correction on the attention risk conduction algorithm of the subsequent graph by using the model weight coefficient, the coefficient is subjected to interval division and is divided into four intervals: the minimum, low, medium and high values are represented by uniform coefficients, and the optimal coefficients obtained by using a grid search method for each interval are shown in the table above.
In step S102, the graph of the enterprise relationship network with the right may be represented by a neighboring matrix with the right, and the neighboring matrix with the right is used as an input of the graph attention network, where the neighboring matrix with the right is:
Figure BDA0003369537050000082
wherein A [ i ] [ j ] is a weighted adjacency matrix, i, j are nodes in the network, E is a set of edges in the network, w is a weight of an edge, and
Figure BDA0003369537050000091
wherein R isijIs a set of relationships between nodes i and j, rijIs the relation weight between the nodes i and j, and n is the relation number between the nodes i and j.
As an example, the weighted adjacency matrix has a size of N × N, where N is the number of enterprise nodes in the network graph.
In an embodiment of the present invention, the state information of the enterprise includes: enterprise basic information, financial indexes and early warning factors. In a specific implementation process, the state information of the enterprise nodes in the network graph may be represented by a matrix X, where the size of the matrix X may be N × F, N is the number of the enterprise nodes in the network, and F is the number of the enterprise state information. In one embodiment of the present invention, a total of 80 pieces of enterprise status information are selected, some of which may be shown in the following table.
Figure BDA0003369537050000092
In the actual implementation process, continuous state information can be directly valued, one-hot coding can be adopted for discrete state information, and all information is spliced into an initial vector
Figure BDA0003369537050000101
As an input.
And finally, outputting a risk prediction matrix Y through a graph attention network (a linear layer-2 graph attention layer-an output layer), wherein the size of Y is 1 x N, and N is the number of enterprise nodes in the network. The output result is a risk level or risk probability.
Graph attention networks (GATs) are a method of processing network data using a deep learning framework and an attention mechanism in a graph representation learning domain. The input of the GAT is the feature vector of the network graph structure and the node, and the output is the node vector representation after the neighbor node information is aggregated. Unlike a classical Graph Convolution Network (GCN), the graph attention network weights and sums up the attention scores of neighboring nodes when aggregating the neighboring node information, that is, the higher the attention of the nodes is, the larger the proportion of the node information is. The method has higher robustness to disturbance, and can be trained without obtaining the information of the whole network structure. After the vector representation of the node update is obtained, tasks such as node classification or link prediction can be performed. The GAT is used as an improved network representation learning method and is widely applied to the fields of community division, node classification, network graph generation and the like.
In the invention, the attention network is adopted to aggregate the associated information and predict the risk state of the enterprise and the associated enterprise. Specifically, an incidence relation network structure (represented by a network weighted adjacency matrix A) is input, state information (numerical value, category or level) of each enterprise node in the network at the time t is input, the input value passes through a graph attention network model architecture (a linear layer-2 graph attention layer-output layer), and discrete risk level or risk probability after softmax at the time t +1 is output.
In the embodiment of the invention, the graph attention network sequentially comprises a linear layer, two graph attention layers and an output layer.
The linear layer adopts MLP, and the linear layer has the function of performing linear mapping on input features F → F'.
The graph attention layer is used for calculating the attention weight of the nodes considering the incidence relation among the enterprises and outputting the node representation. Specifically, in the graph attention layer, the attention weight of a node is calculated by using the following formula:
Figure BDA0003369537050000111
in a conventional graph attention network, the influence of the weight of edges between nodes is ignored in graph attention calculation, so that interpretable content based on services is lost. In the invention, in order to better apply the graph attention network to the enterprise multi-incidence relation network, the attention score calculation method of the graph attention network is optimized, incidence relation weight correction based on business knowledge is added, the model interpretability is enhanced, and a good effect is obtained in an actual data experiment. Specifically, the result of the attention weight matrix is adjusted through the adjacency matrix with the edge weight, financial business knowledge based on the incidence relation among enterprises is fused, and the interpretability of the final prediction result is enhanced.
In the invention, in the process of masking an attention weight matrix (mask), the method of multiplying the weight of an edge and the attention weight is used for sparsifying, a normalized object is restricted to the neighbor of each node, and finally the attention moment matrix of the node is obtained as follows:
Figure BDA0003369537050000112
the updated node representation formula is as follows:
Figure BDA0003369537050000113
the meaning of the parameters referred to in the above formula is as follows:
Figure BDA0003369537050000114
representing the l-th layer attention weight of the nodes i and j;
Figure BDA0003369537050000115
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure BDA0003369537050000116
representing the ith layer feature vector of the node i;
Figure BDA0003369537050000117
representing the ith layer feature vector of the node j;
Figure BDA0003369537050000118
representing the l +1 layer characteristic vector of the node i;
Figure BDA0003369537050000121
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weight of the edges of the nodes i and j;
σ denotes a nonlinear activation function.
The output layer is used to derive a predicted enterprise risk level or probability using softmax.
The graph attention network model uses dropout to prevent overfitting in the whole training process, and is trained through an Adam gradient descent method. The loss function is a Cross Entropy (Cross Entropy) loss function.
In the invention, the association information is aggregated by adopting the graph attention network to predict the risk state of the enterprise and the associated enterprise, and in the calculation of the attention score of the graph attention network, the association relation weight correction based on business knowledge is added, so that the interpretability of the model is enhanced, and a good effect is obtained in an actual data experiment.
Example two
As shown in fig. 2, another aspect of the present invention further includes a functional module architecture completely corresponding to the foregoing method flow, that is, an embodiment of the present invention further provides an enterprise risk prediction apparatus based on a graph attention network, including:
the authorized enterprise incidence relation network graph constructing module 201 is used for constructing an authorized enterprise incidence relation network graph with the relationship among enterprises having weight;
the graph attention network prediction module 202 is configured to take the authorized enterprise association relationship network graph and the state information of each enterprise in the network graph at the previous time as input, and output the risk level or risk probability of the enterprise at the next time through the graph attention network; wherein, in the graph attention network, attention weight is calculated by using the enterprise association relation with the right.
The authorized enterprise incidence relation network graph construction module is specifically configured to:
acquiring enterprise association relation data;
and taking the enterprises as nodes, and taking the incidence relation among the enterprises with the weights as edges to obtain a network graph of the incidence relation of the enterprises with the weights.
Further, in the graph attention network prediction module, the graph of the weighted enterprise incidence relation network is represented by a weighted adjacency matrix, and the weighted adjacency matrix is used as an input of the graph attention network, wherein the weighted adjacency matrix is:
Figure BDA0003369537050000131
wherein A [ i ] [ j ] is a weighted adjacency matrix, i, j are nodes in the network, E is a set of edges in the network, w is a weight of an edge, and
Figure BDA0003369537050000132
wherein R isijIs a set of relationships between nodes i and j, rijIs the relation weight between the nodes i and j, and n is the relation number between the nodes i and j.
And determining the relation weight between the nodes i and j according to the relation between the nodes i and j.
Further, the relationship between the nodes i and j comprises a debt type, an equity type, a supply type and a macro type, and the debt type relationship comprises a guarantee providing relationship, a guaranteed relationship and an inter-insurance joint guarantee; the stock right type relationship comprises a stock control relationship, a stock participation relationship, a relationship controlled by a third enterprise, a common actual controller, a relationship of relatives of the actual controller and a common high-level manager; the supply-type relationship comprises a supply chain; the macro-type relationship comprises the same industry and the same region.
Further, the state information of the enterprise includes: enterprise basic information, financial indexes and early warning factors.
Further, the graph attention network comprises a linear layer, two graph attention layers and an output layer in sequence;
in the graph attention layer, an attention matrix of a node is calculated according to the following formula:
Figure BDA0003369537050000133
Figure BDA0003369537050000134
wherein the content of the first and second substances,
Figure BDA0003369537050000141
representing the l-th layer attention weight of the nodes i and j;
Figure BDA0003369537050000142
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure BDA0003369537050000143
representing the ith layer feature vector of the node i;
Figure BDA0003369537050000144
representing the ith layer feature vector of the node j;
Figure BDA0003369537050000145
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weights of the edges of nodes i and j.
The device can be implemented by the enterprise risk prediction method based on the graph attention network provided in the first embodiment, and specific implementation methods can be referred to the description in the first embodiment and are not described herein again.
The invention also provides a memory storing a plurality of instructions for implementing the method according to the first embodiment.
The invention also provides an electronic device comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, and the instructions can be loaded and executed by the processor to enable the processor to execute the method according to the first embodiment.
While preferred embodiments of the present invention have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all such alterations and modifications as fall within the scope of the invention. It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims (8)

1. An enterprise risk prediction method based on a graph attention network is characterized by comprising the following steps:
constructing a network diagram of the relationship of the enterprises with the weight and the association relationship of the enterprises with the weight;
taking the network diagram of the association relationship of the authorized enterprises and the state information of each enterprise in the network diagram at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the network of attention by the diagram; wherein, in the graph attention network, the attention weight of the node is calculated by using the association relation of the enterprise with the weight;
in the network diagram of the incidence relation of the enterprises with the weight, the enterprises are used as nodes, and the incidence relation between the enterprises with the weight is used as an edge; the weight of the edge between the nodes i and j is determined by expert fitting according to the incidence relation between the nodes i and j and based on financial business knowledge;
the graph attention network sequentially comprises a linear layer, two graph attention layers and an output layer;
in the graph attention layer, the graph attention layer is thinned by using a mode of multiplying the weight of the edge and the attention weight, and the attention matrix of the node is calculated according to the following formula:
Figure FDA0003369537040000011
Figure FDA0003369537040000012
wherein the content of the first and second substances,
Figure FDA0003369537040000013
representing the l-th layer attention weight of the nodes i and j;
Figure FDA0003369537040000014
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure FDA0003369537040000015
representing the ith layer feature vector of the node i;
Figure FDA0003369537040000016
representing the ith layer feature vector of the node j;
Figure FDA0003369537040000017
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weights of the edges of nodes i and j.
2. The graph attention network-based enterprise risk prediction method of claim 1, wherein the constructing a weighted enterprise incidence relation network graph with weighted inter-enterprise relations comprises:
acquiring enterprise association relation data;
and taking the enterprises as nodes, and taking the incidence relation among the enterprises with the weights as edges to obtain a network graph of the incidence relation of the enterprises with the weights.
3. The graph attention network based enterprise risk prediction method of claim 1, wherein the graph of the weighted enterprise incidence relation network is represented by a weighted adjacency matrix, and the weighted adjacency matrix is used as an input of the graph attention network, wherein the weighted adjacency matrix is:
Figure FDA0003369537040000021
wherein A [ i ] [ j ] is a weighted adjacency matrix, i, j are nodes in the network, E is a set of edges in the network, w is a weight of an edge, and
Figure FDA0003369537040000022
wherein R isijIs a set of relationships between nodes i and j, rijIs the relation weight between the nodes i and j, and n is the relation number between the nodes i and j.
4. The graph attention network based enterprise risk prediction method of claim 1 wherein the relationships between nodes i and j include debt type, equity type, supply type and macro type, the debt type relationships include offering a guaranty relationship, an inter-guaranty joint guaranty; the stock right type relationship comprises a stock control relationship, a stock participation relationship, a relationship controlled by a third enterprise, a common actual controller, a relationship of relatives of the actual controller and a common high-level manager; the supply-type relationship comprises a supply chain; the macro-type relationship comprises the same industry and the same region.
5. The graph attention network based enterprise risk prediction method of claim 1, wherein the state information of the enterprise comprises: enterprise basic information, financial indexes and early warning factors.
6. An enterprise risk prediction device based on a graph attention network, comprising:
the system comprises a weighted enterprise incidence relation network graph construction module, a weighted enterprise incidence relation network graph generation module and a weighted enterprise incidence relation network graph generation module, wherein the weighted enterprise incidence relation network graph construction module is used for constructing a weighted enterprise incidence relation network graph with inter-enterprise relations; in the network diagram of the incidence relation of the enterprises with the weight, the enterprises are used as nodes, and the incidence relation between the enterprises with the weight is used as an edge; the weight of the edge between the nodes i and j is determined by expert fitting according to the incidence relation between the nodes i and j and based on financial business knowledge;
the graph attention network prediction module is used for taking the authorized enterprise incidence relation network graph and the state information of each enterprise in the network graph at the previous moment as input, and outputting the risk level or risk probability of the enterprise at the next moment through the graph attention network; wherein, in the graph attention network, the attention weight is calculated by using the association relation of the enterprise with the weight; the graph attention network sequentially comprises a linear layer, two graph attention layers and an output layer;
in the graph attention layer, the graph attention layer is thinned by using a mode of multiplying the weight of the edge and the attention weight, and the attention matrix of the node is calculated according to the following formula:
Figure FDA0003369537040000031
Figure FDA0003369537040000032
wherein the content of the first and second substances,
Figure FDA0003369537040000033
representing the l-th layer attention weight of the nodes i and j;
Figure FDA0003369537040000034
representing the l-th layer weight vector;
W(l)representing the characteristic transformation weight parameter of the l-th layer node;
Figure FDA0003369537040000035
representing the ith layer feature vector of the node i;
Figure FDA0003369537040000036
representing the ith layer feature vector of the node j;
Figure FDA0003369537040000041
representing attention weight of the nodes i and j after the association relation weight correction;
wijrepresenting the weights of the edges of nodes i and j.
7. A memory storing a plurality of instructions for implementing the method of any one of claims 1-5.
8. An electronic device comprising a processor and a memory coupled to the processor, the memory storing a plurality of instructions that are loadable and executable by the processor to enable the processor to perform the method according to any of claims 1-5.
CN202111394584.1A 2021-11-23 2021-11-23 Enterprise risk prediction method and device based on graph attention network and electronic equipment Active CN114066081B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111394584.1A CN114066081B (en) 2021-11-23 2021-11-23 Enterprise risk prediction method and device based on graph attention network and electronic equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111394584.1A CN114066081B (en) 2021-11-23 2021-11-23 Enterprise risk prediction method and device based on graph attention network and electronic equipment

Publications (2)

Publication Number Publication Date
CN114066081A CN114066081A (en) 2022-02-18
CN114066081B true CN114066081B (en) 2022-04-26

Family

ID=80279436

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111394584.1A Active CN114066081B (en) 2021-11-23 2021-11-23 Enterprise risk prediction method and device based on graph attention network and electronic equipment

Country Status (1)

Country Link
CN (1) CN114066081B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116151635B (en) * 2023-04-19 2024-03-08 深圳市迪博企业风险管理技术有限公司 Optimization method and device for decision-making of anti-risk enterprises based on multidimensional relation graph

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110570111A (en) * 2019-08-30 2019-12-13 阿里巴巴集团控股有限公司 Enterprise risk prediction method, model training method, device and equipment
CN111161535A (en) * 2019-12-23 2020-05-15 山东大学 Attention mechanism-based graph neural network traffic flow prediction method and system
CN113361962A (en) * 2021-06-30 2021-09-07 支付宝(杭州)信息技术有限公司 Method and device for identifying enterprise risk based on block chain network
WO2021179838A1 (en) * 2020-03-10 2021-09-16 支付宝(杭州)信息技术有限公司 Prediction method and system based on heterogeneous graph neural network model

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019220128A1 (en) * 2018-05-18 2019-11-21 Benevolentai Technology Limited Graph neutral networks with attention

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110570111A (en) * 2019-08-30 2019-12-13 阿里巴巴集团控股有限公司 Enterprise risk prediction method, model training method, device and equipment
CN111161535A (en) * 2019-12-23 2020-05-15 山东大学 Attention mechanism-based graph neural network traffic flow prediction method and system
WO2021179838A1 (en) * 2020-03-10 2021-09-16 支付宝(杭州)信息技术有限公司 Prediction method and system based on heterogeneous graph neural network model
CN113361962A (en) * 2021-06-30 2021-09-07 支付宝(杭州)信息技术有限公司 Method and device for identifying enterprise risk based on block chain network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于自注意力模型的企业关系抽取;张豪杰等;《电子测量技术》;20200614(第10期);101-105页 *

Also Published As

Publication number Publication date
CN114066081A (en) 2022-02-18

Similar Documents

Publication Publication Date Title
Lagergren et al. Learning partial differential equations for biological transport models from noisy spatio-temporal data
Pulido et al. Particle swarm optimization of ensemble neural networks with fuzzy aggregation for time series prediction of the Mexican Stock Exchange
Meinshausen et al. Monte Carlo methods for the valuation of multiple‐exercise options
CN109902222A (en) Recommendation method and device
Liu et al. Predicting housing price in China based on long short-term memory incorporating modified genetic algorithm
Aydin et al. Prediction of financial crisis with artificial neural network: an empirical analysis on Turkey
CN111695719A (en) User value prediction method and system
CN109766454A (en) A kind of investor's classification method, device, equipment and medium
Bastos et al. Explainable models of credit losses
Adam et al. Forecasting of peak electricity demand in Mauritius using the non-homogeneous Gompertz diffusion process
Misund Financial ratios and prediction on corporate bankruptcy in the Atlantic salmon industry
Kultur et al. Ensemble of neural networks with associative memory (ENNA) for estimating software development costs
CN107563542A (en) Data predication method and device and electronic equipment
Gharleghi et al. Predicting exchange rates using a novel “cointegration based neuro-fuzzy system”
CN114066081B (en) Enterprise risk prediction method and device based on graph attention network and electronic equipment
Zhong et al. Dynamically evolving deep neural networks with continuous online learning
Sharifi et al. Banks credit risk prediction with optimized ANN based on improved owl search algorithm
Kasgari et al. Price manipulation fraud detection by Intelligent Visual Fraud surveillance system
CN109741172B (en) Credit early warning method, device, system and storage medium
Fazla et al. Joint optimization of linear and nonlinear models for sequential regression
Balcilar et al. Was the recent downturn in US real GDP predictable?
Bastos et al. Nonparametric models of financial leverage decisions
Chen et al. An Optimized BP Neural Network Model and Its Application in the Credit Evaluation of Venture Loans
US20220138552A1 (en) Adapting ai models from one domain to another
CN115063145A (en) Transaction risk factor prediction method and device, electronic equipment and storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant