A method for constructing an enterprise risk assessment model, an assessment method and a device

Abstract

The present invention provides a method for constructing an enterprise risk assessment model, an assessment method, and a device, which relate to the field of artificial intelligence technology. The method includes: constructing an investment relationship network of a target enterprise, a syndicate relationship network of venture capital, and a cooperation relationship network of the target enterprise; constructing a first investment network diagram of the target enterprise, a first patent cooperation network diagram, and a first syndicate network diagram of the venture capital associated with the target enterprise, and generating a first heterogeneous network. Repeat the above steps to construct investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams at different times. Generate a heterogeneous network atlas according to the three; train a prediction analysis model according to the heterogeneous network atlas to obtain an enterprise risk assessment model. By clarifying the cooperation relationship between an enterprise and other enterprises and institutions, it can assist the development of the network relationship of start-up enterprises, help venture capital companies screen high-quality start-up enterprises, and optimize the allocation of resources for higher returns.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular, to a method for constructing an enterprise risk assessment model, an assessment method, and a device. Background Art

[0002] Currently, the relationship network among enterprises plays a crucial role in the future development of start-up enterprises. However, most current prediction models rely on using unstructured data, such as industry location and company size, while ignoring the structured data of the relationship network among enterprises. At the same time, the results of current homogeneous graph network prediction have problems of information loss and underfitting of the relationship network, resulting in a reduced mutual correlation between different types of networks. Summary of the Invention

[0003] The problem to be solved by the present invention is how to construct a heterogeneous relationship network and then evaluate the risks of start-up enterprises to assist start-up enterprises in understanding the risks of company development.

[0004] To solve the above problems, the present invention provides a method for constructing an enterprise risk assessment model, including:

[0005] Step S1, obtaining investment event data before a preset moment and constructing an investment relationship network of a target enterprise;

[0006] Step S2, constructing a syndicate relationship network of the venture capital according to the venture capital in the investment relationship network;

[0007] Step S3, constructing a cooperation relationship network of the target enterprise according to the cooperation enterprises in the investment relationship network;

[0008] Step S4, constructing a first investment network diagram, a first patent cooperation network diagram of the target enterprise, and a first syndicate network diagram of the venture capital according to the investment relationship network, the syndicate relationship network, and the cooperation relationship network within a first preset time interval, and generating a first heterogeneous network;

[0009] Step S5, constructing a second investment network diagram, a second patent cooperation network diagram of the target enterprise, and a second syndicate network diagram of the venture capital according to the investment relationship network, the syndicate relationship network, and the cooperation relationship network within a second preset time interval, and generating a second heterogeneous network;

[0010] Step S6: Repeat steps S4 and S5. Based on the investment relationship network, syndicate relationship network, and cooperation relationship network within different preset time intervals, construct multiple corresponding investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams. Generate a heterogeneous network atlas according to all the investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams.

[0011] Step S7: Train a preset predictive analysis model according to the heterogeneous network atlas to obtain an enterprise risk assessment model.

[0012] Compared with the prior art: By collecting investment event data before a preset moment of a target enterprise, the present invention can construct an investment relationship network of the target enterprise, which clarifies the investment relationship between the target enterprise and the investors; in the investment relationship network, identify and extract the investment relationships involving venture capital, and then construct a syndicate relationship network of venture capital to represent the investment relationship between the target enterprise and venture capital; in the investment relationship network, identify and extract the enterprises having a cooperation relationship with the target enterprise, and then construct a cooperation relationship network of the target enterprise to represent the cooperation relationship between the target enterprise and its partners; within the selected preset time interval, use the investment relationship network, syndicate relationship network, and cooperation relationship network to construct a first investment network diagram of the target enterprise, a first patent cooperation network diagram, and a first syndicate network diagram of venture capital, and then combine these networks into a first heterogeneous network for further analysis and prediction. By repeating these steps, heterogeneous network diagrams within different time intervals can be obtained, a heterogeneous network atlas can be generated, and the predictive analysis model can be trained. This model can predict the business operation risks of the target enterprise. Among them, historical data can be used to train the prediction model to improve its prediction accuracy and reliability, and finally an enterprise risk assessment model is obtained. The present invention analyzes and evaluates the investment relationship network and cooperation relationship network of the target enterprise, combines the predictive analysis model, clarifies the corresponding information on the current situation and potential of the enterprise, helps start-up enterprises and investors to clarify the current development situation of the enterprise, evaluate the future development risks, assist in the network relationship development of start-up enterprises by evaluating the importance of affiliated companies, guide them to establish key relationships, and can also help venture capital companies screen high-quality start-up enterprises and optimize the allocation of resources for higher returns.

[0013] Optionally, the obtaining of the investment event data before the preset moment, constructing the investment relationship network of the target enterprise; constructing the syndicate relationship network of the venture capital according to the venture capital in the investment relationship network; constructing the cooperation relationship network of the target enterprise according to the cooperative enterprises in the investment relationship network includes:

[0014] Generate a characteristic matrix of the industry to which the target enterprise belongs.

[0015] Generate a characteristic matrix of the location of the target enterprise according to the location of the target enterprise;

[0016] Generate a characteristic matrix of the location of the venture capital according to the location of the venture capital;

[0017] Generate a category characteristic matrix according to the category of the venture capital;

[0018] Generate a characteristic matrix of the location of the cooperative enterprise according to the location of the cooperative enterprise;

[0019] Obtain a complete node characteristic matrix according to the industry characteristic matrix, the characteristic matrix of the location of the target enterprise, the characteristic matrix of the location of the venture capital, the category characteristic matrix, and the characteristic matrix of the location of the cooperative enterprise;

[0020] Label the target enterprise, the venture capital, and the cooperative enterprise to obtain a node label matrix;

[0021] Construct the investment relationship network, the syndicate relationship network, and the cooperative relationship network according to the complete node characteristic matrix and the node label matrix.

[0022] Optionally, the prediction analysis model includes a network structure composed of a heterogeneous node attention layer, a heterogeneous meta-path attention layer, and a sequence representation layer. Training the preset prediction analysis model according to the heterogeneous network atlas to obtain an enterprise risk assessment model includes:

[0023] Step A1, input the heterogeneous network atlas into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information, and cooperative relationship network information;

[0024] Step A2, input the investment relationship network information, the syndicate relationship network information, and the cooperative relationship network information into the heterogeneous meta-path attention layer, where the heterogeneous meta-path attention layer includes a hybrid layer module and a meta-path attention module;

[0025] Step A21, initially combine the investment relationship network information and the syndicate relationship network information through the hybrid layer module to obtain fused network information;

[0026] Step A22, input the fused network information and the cooperative relationship network information into the meta-path attention module, obtain meta-path weights and a node attention matrix, and fuse the fused network information and the cooperative relationship network information to obtain node embeddings;

[0027] Step A3: Obtain all the time series in the heterogeneous network atlas through the sequence characterization layer. Repeat steps A1 - A2 to obtain the node embeddings corresponding to all the time series. Make predictions for each node embedding to obtain temporary prediction results. Optimize the preset prediction analysis model according to the prediction results and the heterogeneous network atlas to obtain the enterprise risk assessment model.

[0028] Optionally, the step of inputting the heterogeneous graph neural network into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information includes:

[0029] Step A11: Obtain investment relationship meta - paths, syndicate relationship meta - paths, and cooperation relationship meta - paths based on the investment relationship network, the syndicate relationship network, and the cooperation relationship network.

[0030] Step A12: Obtain the node weight of the first node to the second node in the investment relationship meta - path, the syndicate relationship meta - path, and the cooperation relationship meta - path according to the node weight formula.

[0031] Among them, the node weight formula is:

[0032]

[0033] Among them, is the node weight, is the node - level attention vector under the meta - path Φ p σ is the activation function, || represents the concatenation operation, h i represents the feature vector of the i - th node, h j represents the feature vector of the j - th node.

[0034] Step A13: Standardize the node weights according to the standardization formula to obtain attention weights.

[0035] Among them, the standardization formula is:

[0036]

[0037] Among them, is the attention weight, is the node weight.

[0038] Step A14: Add up the influence weights of all adjacent nodes of the first node, and repeat steps A12 and A13 multiple times to obtain multiple attention weights. Concatenate the multiple attention weights to obtain the node attention matrix of the first node under the investment relationship meta-path, the syndicate relationship meta-path, and the cooperation relationship meta-path. Add up the influence weights of all adjacent nodes of each node, repeat steps A12 to A14 multiple times, and combine to obtain the complete node attention matrix;

[0039] Step A15: Obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information based on the complete node attention matrix.

[0040] Optionally, the initial merging of the investment relationship network information and the syndicate relationship network information by the hybrid layer module to obtain fusion network information includes:

[0041] Step A211: Concatenate the investment relationship network information and the syndicate relationship network information through a linear fusion formula and input them into a linear function layer to obtain a linear fusion matrix, and obtain fusion network information based on the linear fusion matrix.

[0042] Among them, the linear fusion formula is:

[0043]

[0044] Among them, is the linear fusion matrix, and || represents the concatenation operation.

[0045] Optionally, inputting the fusion network information and the cooperation relationship network information into the meta-path attention module to obtain meta-path weights and a node attention matrix, and fusing the fusion network information and the cooperation relationship network information to obtain node embeddings includes:

[0046] Step A221: Process the fusion network information and the cooperation relationship network information through a meta-path weight formula to obtain the meta-path weights.

[0047] Among them, the meta-path weight formula is:

[0048]

[0049] Among them, is the meta-path weight, D is the weight matrix, q T is the semantic attention vector, and b is the bias;

[0050] Step A222: Standardize the meta-path weights through the node attention matrix formula to obtain the node attention matrix.

[0051] Among them, the node attention matrix formula is:

[0052]

[0053] Among them, is the influence weight, is the meta-path weight;

[0054] Step A223: Process the meta-path weight and the node attention matrix through the node embedding formula to obtain the node embedding.

[0055] Among them, the node embedding formula is:

[0056]

[0057] Among them, is the meta-path weight, is the influence weight, and M is the node embedding.

[0058] Optionally, predicting each of the node embeddings to obtain a temporary prediction result includes:

[0059] Predict each node embedding through the prediction processing formula to obtain the temporary prediction result.

[0060] Among them, the prediction processing formula is:

[0061] U = LSTM(M t-k , M t-k+1 ,..., M t ),

[0062] Among them, U is the temporary prediction result, M t-k is the (t - k)-th node embedding, M t-k+1 is the (t - k + 1)-th node embedding, and M t is the k-th node embedding.

[0063] To solve the above problems, the present invention also provides an enterprise risk assessment model construction device, which is characterized by including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement any one of the enterprise risk assessment model construction methods when executing the computer program.

[0064] The advantages of the startup enterprise risk assessment device and the enterprise risk assessment model construction method of the present invention over the prior art are the same and will not be elaborated here.

[0065] To solve the above problems, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for constructing an enterprise risk assessment model described in any one of the above is implemented.

[0066] The advantages of the computer-readable storage medium of the present invention and the method for constructing an enterprise risk assessment model of the present invention are the same as those of the prior art, and will not be elaborated here.

[0067] To solve the above problems, the present invention also provides an enterprise risk assessment method, including:

[0068] Obtain a heterogeneous network map set of a startup enterprise;

[0069] Input the heterogeneous network map set into the enterprise risk assessment model to obtain an enterprise risk assessment report.

[0070] The advantages of the enterprise risk assessment method of the present invention and the method for constructing an enterprise risk assessment model of the present invention are the same as those of the prior art, and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a schematic flowchart of the method for constructing an enterprise risk assessment model in an embodiment of the present invention;

[0072] Figure 2 It is an internal structure diagram of a computer device in an embodiment of the present invention;

[0073] Figure 3 A schematic structural diagram of an enterprise risk assessment model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0074] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be given with reference to the accompanying drawings.

[0075] To solve the above problems, in combination with Figure 1 as shown, the present invention provides a method for constructing an enterprise risk assessment model, including:

[0076] Step S1, obtain investment event data before a preset time, and construct an investment relationship network of a target enterprise;

[0077] Specifically, the investment relationship network in this embodiment is a graphical structure, where each node represents an enterprise, and each edge represents the investment relationship between two enterprises. Among them, the degree of a node represents the number of investment activities of the enterprise or the number of times the enterprise receives investment, and the weight of an edge can represent the size of the investment amount or the intensity of the investment relationship. By constructing and analyzing the investment relationship network of the target enterprise, it can help to understand the associations between enterprises, the flow of funds, and the industrial layout, etc.

[0078] Step S2: Construct the syndicate relationship network of the venture capital based on the venture capital in the investment relationship network;

[0079] Specifically, the syndicate relationship network in this embodiment mainly refers to a network model. In this network, the weight of an edge is determined by the sum of the degrees of the connected nodes. The syndicate relationship network has certain applications in studying the interactions and connection intensities between nodes in the network. By analyzing the edge weights in the network, it reveals the connection patterns and importance between nodes, which helps to understand the structure and function of complex systems. Based on the heterogeneous graph neural network, the cooperation relationship network of the target enterprise can be constructed according to the cooperating enterprises in the investment relationship network. In this cooperation relationship network, each node represents an enterprise, and the edge represents the cooperation relationship between two enterprises.

[0080] Step S3: Construct the cooperation relationship network of the target enterprise according to the cooperating enterprises in the investment relationship network;

[0081] Specifically, in the cooperation relationship network of this embodiment, the degree of a node represents the number of times an enterprise cooperates with other enterprises, and the weight of an edge represents the intensity of the cooperation relationship or the scale of cooperation. By constructing the cooperation relationship network of the target enterprise, it can help to understand the partners of the target enterprise, the density of the cooperation relationship, and its position in the entire industrial chain. Through the method of heterogeneous graph neural network, these complex cooperation relationship networks can be better mined and analyzed.

[0082] Specifically, first, collect the investment event data of the target enterprise before the preset time, including information such as the investor and the investment time. Then, use this data to construct the investment relationship network of the target enterprise. In this network, the target enterprise serves as the central node, and there is an investment edge between the target enterprise and the investor. The edge of this network is 0 / 1. If the edge is 1, then there is an investment edge between the target enterprise and the investor. Conversely, if the value of the edge is 0, it means that there is no investment edge between the two; Based on the investment relationship network, identify and extract the investment relationships involving venture capital. Among them, venture capital can be a venture capital firm, a private equity firm, or other investors. Then, construct the syndicate relationship network of venture capital according to these investment relationships. In this network, venture capital serves as a node, and there is an edge between two venture capitals. In this network, the edge indicates whether there is cooperation between two venture capitals. If there is, the corresponding value of the edge is 1; Based on the investment relationship network, identify and extract the enterprises that have a cooperative relationship with the target enterprise. These cooperative enterprises can be suppliers, partners, customers, or other related enterprises. Then, construct the cooperative relationship network of the target enterprise according to these cooperative relationships. In this network, the target enterprise and the cooperative enterprises serve as nodes, and there is a cooperative edge between two nodes. In this network, the edge indicates whether there is cooperation between the target enterprise and the cooperative enterprise. If there is, the corresponding value of the edge is 1. Through this embodiment, detailed information about the investment and cooperative relationships of the target enterprise can be provided, assisting the enterprise in understanding the current development risks of the company and providing a data basis for subsequent prediction of the enterprise's development risks.

[0083] Step S4, according to the investment relationship network, the syndicate relationship network, and the cooperative relationship network within the first preset time interval, construct the first investment network diagram, the first patent cooperation network diagram of the target enterprise, and the first syndicate network diagram of the venture capital, and generate the first heterogeneous network;

[0084] Step S5, according to the investment relationship network, the syndicate relationship network, and the cooperative relationship network within the second preset time interval, construct the second investment network diagram, the second patent cooperation network diagram of the target enterprise, and the second syndicate network diagram of the venture capital, and generate the second heterogeneous network;

[0085] Specifically, the first preset time interval and the second preset time interval in this embodiment are consecutive preset time intervals. It can be understood that their time values are set artificially, and the first and the second are only used to distinguish the front and back time sequences.

[0086] Step S6: Repeat steps S4 and S5. Based on the investment relationship network, syndicate relationship network, and cooperation relationship network within different preset time intervals, construct multiple corresponding investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams, and generate a heterogeneous network atlas according to all the investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams.

[0087] Specifically, adopt the method of a sliding time window. The length of each sliding window is T interval , and the sliding step size is T step . A total of N time windows T are set n . According to the relationships occurring within the time period [T current _T interval , T current , construct the investment network diagram of the startup Patent cooperation network diagram and the syndicate network diagram of venture capital VC Thus, generate the heterogeneous network G i . Then, according to the relationships occurring within the time period [T current _T step _T interval , T current _T step , construct the investment network diagram of the startup Patent cooperation network diagram and the syndicate network diagram of VC Thus, generate the heterogeneous network G i_1 . And so on, respectively construct the investment network diagrams, patent cooperation network diagrams, and syndicate network diagrams of the remaining N_2 startups. Eventually, N network diagrams will be obtained for each type of relationship, forming N heterogeneous network diagrams, and further obtaining a heterogeneous network atlas.

[0088] Step S7: Train the prediction analysis model according to the heterogeneous network atlas to obtain an enterprise risk assessment model.

[0089] Specifically, arrange N time periods into the training set, validation set, and test set in chronological order. Use the network data in the training set and validation set to train and learn the prediction analysis model to obtain an enterprise risk assessment model; put the network data in the test set into the model for learning, and analyze and evaluate the important network relationships and partnership relationships of start-up enterprises through the results of the model. Among them, the prediction analysis model in this embodiment adopts a heterogeneous graph neural network structure, which is a deep learning model for processing complex graph data with multiple types of nodes and edges. In this embodiment, a heterogeneous network atlas is constructed using the investment relationship network of the target enterprise, the syndicate relationship network of venture capital, and the cooperation relationship network, and the prediction analysis model is trained using the heterogeneous network atlas. At the same time, we can also use the investment relationship network, the syndicate relationship network of venture capital, and the cooperation relationship network for more in-depth analysis and evaluation. For example, evaluate indicators such as the stability and influence of the investment relationship and cooperation relationship of the target enterprise. Finally, based on these analysis and evaluation results, generate a risk assessment report to provide insights and suggestions on the future financing situation, investment relationship, and cooperation relationship of the target enterprise.

[0090] Compared with the prior art: In this embodiment, by collecting investment event data of a target enterprise before a preset time, an investment relationship network of the target enterprise can be constructed, which clarifies the investment relationship between the target enterprise and the investors; based on the investment relationship network, the investment relationships involving venture capital are identified and extracted, and then a syndicate relationship network of venture capital is constructed to represent the investment relationship between the target enterprise and venture capital; based on the investment relationship network, the enterprises having a cooperation relationship with the target enterprise are identified and extracted, and then a cooperation relationship network of the target enterprise is constructed to represent the cooperation relationship between the target enterprise and its partners; based on the investment relationship network, the syndicate relationship network and the cooperation relationship network, further fusion is carried out, that is, within a selected preset time interval, the investment relationship network, the syndicate relationship network and the cooperation relationship network are used to construct a first investment network diagram of the target enterprise, a first patent cooperation network diagram and a first syndicate network diagram of venture capital, and then these networks are combined into a first heterogeneous network for further analysis and prediction. By repeating these steps, heterogeneous network diagrams within different time intervals can be obtained, a heterogeneous network atlas can be generated, and a prediction analysis model can be trained. It can be understood that compared with the blind network merging or single network merging of the prior art, in this embodiment, the investment relationship network and the syndicate relationship network need to be merged first, and then fused with the cooperation relationship network. The network merging method of this embodiment takes into account the information loss caused by single network merging, and realizes better network fusion by considering the network merging order, that is, considering the relationships and interactions between networks. This model can predict the business operation risks of the target enterprise. Among them, historical data can also be used to train the prediction model to improve its prediction accuracy and reliability, and finally an enterprise risk assessment model is obtained. In this embodiment, the investment relationship network and the cooperation relationship network of the target enterprise are analyzed and evaluated, combined with the prediction analysis model, to clarify the corresponding information of the enterprise's current situation and potential, help start-up enterprises and investors to clarify the current development situation of the enterprise, evaluate the future development risks, assist the network relationship development of start-up enterprises by evaluating the importance of associated companies, guide them to establish key relationships, and can also help venture capital companies screen high-quality start-up enterprises and optimize resource allocation for higher returns.

[0091] In one embodiment, in combination with Figure 3As shown, the circle represents the startup, the triangle represents the venture capital, and the square represents the patent cooperation company. Taking the above three parties as an example, the network relationships corresponding to the startup, venture capital, and patent cooperation company are input into the Heterogeneous node attention layer for independent learning. Among them, the three parties respectively perform multi-head attention learning, then perform meta-path representation, and finally independently output the network information of the investment relationship network, syndicate network, and cooperation relationship network; then the three types of network information learned are input into the Heterogeneous meta-path attention layer. This layer can be divided into two modules: the Hybrid composition Layer module and the meta-path attention module. First, the Hybrid composition Layer module merges the investment relationship network and the syndicate network for the first time to obtain the initially fused network information, and then inputs this part of the initially fused network information and the cooperation relationship network information into the meta-path attention module. By assigning different weights to the network information, the initially fused network information and the network information of the cooperation relationship network are finally fused. Each node in the final network will obtain an embedding representation; for heterogeneous networks at different times, repeat the above steps for learning to obtain the information representation of the same node at different times. The network information representations at different times are input into the Sequential representation layer in chronological order, and the Long Short-Term Memory (LSTM) neural network is used to fuse the temporal information of the nodes to obtain the embedding representation F of the node at time t. F represents the risk probability of the startup's development after time t. The larger the value of F, the lower the risk probability of development.

[0092] Specifically, in the Heterogeneous meta path attention layer part, this embodiment attempts three model variants:

[0093] Model variant 1: There is no Hybrid composition layer, that is, after receiving the three network information from the previous layer, it is directly passed into the meta-path attention layer.

[0094] Model variant 2: There is a Hybrid composition layer, but it is used to merge the investment network and the patent cooperation network for the first time.

[0095] Model variant 3( Figure 3 As shown): There is a Hybrid composition layer, but it is used to merge the investment network and the syndicate network for the first time.

[0096] The results show that the average accuracy of model variant 1 is only 48.60%; the accuracy of model variant 2 is slightly higher than that of variant 1, at 50.93%; the accuracy of model variant 3 is the highest, at approximately 53.20%.

[0097] Optionally, obtaining investment event data before a preset moment, constructing an investment relationship network of a target enterprise; constructing a syndicate relationship network of the venture capital according to the venture capital in the investment relationship network; constructing a cooperation relationship network of the target enterprise according to the cooperation enterprises in the investment relationship network, including:

[0098] Generating an industry feature matrix according to the industry to which the target enterprise belongs;

[0099] Generating a location feature matrix of the target enterprise according to the location of the target enterprise;

[0100] Generating a location feature matrix of the venture capital according to the location of the venture capital;

[0101] Generating a type feature matrix according to the type of the venture capital;

[0102] Generating a location feature matrix of the cooperation enterprise according to the location of the cooperation enterprise;

[0103] Obtaining a complete node feature matrix according to the industry feature matrix, the location feature matrix of the target enterprise, the location feature matrix of the venture capital, the type feature matrix, and the location feature matrix of the cooperation enterprise;

[0104] Labeling the target enterprise, the venture capital, and the cooperation enterprise to obtain a node label matrix;

[0105] Constructing the investment relationship network, the syndicate relationship network, and the cooperation relationship network according to the complete node feature matrix and the node label matrix.

[0106] Specifically, in this embodiment, the target enterprise (start-up enterprise), venture capital VC, and cooperation enterprise are three types of nodes. For these three types of nodes, node feature matrices are respectively generated. Among them, the feature matrix can be generated using One-hot encoding. One-hot encoding, also known as one-hot encoding or one-hot effective encoding. The method is to use an N-bit status register to encode N states. Each state has its own independent register bit, and at any time, only one bit is valid. The node features are merged to obtain a complete node feature matrix H nodes , where each node n i corresponds to a feature vector h i, label the nodes of startups. Assume that the cut-off time of the window is T end . If a startup has a financing event during the time period T end and has the next financing event after time T end , then the startup sets the label to 1, and the startup that does not have the next financing sets the label to 0. Finally, for T n time periods, a node label matrix will be obtained below

[0107] Optionally, the prediction analysis model includes a network structure composed of a heterogeneous node attention layer, a heterogeneous metapath attention layer, and a sequence representation layer. The enterprise risk assessment model obtained by training the preset prediction analysis model according to the heterogeneous network atlas includes:

[0108] Step A1, input the heterogeneous network atlas into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information;

[0109] Step A2, input the investment relationship network information, the syndicate relationship network information, and the cooperation relationship network information into the heterogeneous metapath attention layer, where the heterogeneous metapath attention layer includes a hybrid layer module and a metapath attention module;

[0110] Step A21, initially merge the investment relationship network information and the syndicate relationship network information through the hybrid layer module to obtain fused network information;

[0111] Step A22, input the fused network information and the cooperation relationship network information into the metapath attention module, obtain metapath weights and a node attention matrix, and fuse the fused network information and the cooperation relationship network information to obtain node embeddings;

[0112] Step A3, obtain all time series in the heterogeneous network atlas through the sequence representation layer, repeat steps A1 - A2, obtain the node embeddings corresponding to all the time series, predict each node embedding to obtain a temporary prediction result, and optimize the preset prediction analysis model according to the prediction result and the heterogeneous network atlas to obtain the enterprise risk assessment model.

[0113] Specifically, first, in the heterogeneous node attention part, the information of the three relationship networks is separately learned through the method of the graph attention network (GAT). Subsequently, these three relationship information are passed into the heterogeneous meta path attention part for heterogeneous information merging. Finally, in the sequential representation part, we use the method of the long short-term memory neural network (LSTM) to capture the temporal information of the information. In order to fit the information flow between different networks, a hybrid composition layer is designed in the heterogeneous meta path attention part of the model in this embodiment. Through this layer, selective integration of networks can be achieved, improving the prediction accuracy. During the model construction process, in order to consider the comprehensive effect between different networks, a self-developed hybrid composition layer is introduced. Compared with the model without this layer, after adding this layer, the average prediction accuracy is increased by 4.6%. The average accuracy of the final model is 13.2% higher than that of the manually selected model and is also better than most baseline models.

[0114] Optionally, the step of inputting the heterogeneous graph neural network into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information includes:

[0115] Step A11: Obtain an investment relationship meta-path, a syndicate relationship meta-path, and a cooperation relationship meta-path according to the investment relationship network, the syndicate relationship network, and the cooperation relationship network;

[0116] Step A12: Obtain the node weight of the first node to the second node in the investment relationship meta-path, the syndicate relationship meta-path, and the cooperation relationship meta-path according to the node weight formula,

[0117] where the node weight formula is:

[0118]

[0119] where, is the node weight, is the node-level attention vector under the meta-path Φ p σ is the activation function, || represents the concatenation operation, h i represents the feature vector of the i-th node, h jDenote the feature vector of the j-th node;

[0120] Step A13, according to the normalization formula, normalize the node weights to obtain attention weights,

[0121] where the normalization formula is:

[0122]

[0123] where, is the attention weight, is the node weight;

[0124] Step A14, add up the influence weights of all adjacent nodes of the first node, and repeat Step A12 and Step A13 multiple times to obtain multiple attention weights. Concatenate the multiple attention weights to obtain the node attention matrix of the first node under the investment relationship meta-path, the syndicate relationship meta-path, and the cooperation relationship meta-path. Add up the influence weights of all adjacent nodes of each node, repeat Step A12 to Step A14 multiple times, and combine to obtain the complete node attention matrix;

[0125] Step A15, obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information according to the complete node attention matrix.

[0126] Specifically, the heterogeneous node attention layer is a layer of a neural network used for adaptive feature aggregation and weight assignment for different types of nodes in a heterogeneous graph. By independently learning the attention weights of each node, the heterogeneous node attention layer can perform weighted fusion on the features of each node type to capture the importance and interactions of different node types. The input layer takes the node feature matrices of the investment relationship network, syndicate relationship network, and cooperation relationship network as inputs; the heterogeneous node attention layer applies independent heterogeneous node attention layers to the node feature matrices of the investment relationship network, syndicate relationship network, and cooperation relationship network respectively to learn the attention weights of each node type, performs weighted fusion of the attention weights of each node type with the corresponding node feature matrices to obtain investment relationship network information, syndicate relationship network information, and cooperation relationship network information, and takes the network information after feature aggregation as input to be further processed and analyzed by subsequent layers of the neural network. By independently learning the heterogeneous node attention layer, the network can better understand and utilize the relationships between different types of nodes. For example, the information of the investment relationship network can help capture the connections and influences between the target enterprise and investors, the information of the syndicate relationship network can reflect the cooperation relationships and industry experiences of venture capitals, and the information of the cooperation relationship network can show the cooperation relationships and their influences between the target enterprise and other enterprises. Using the heterogeneous graph neural network and the independently learned heterogeneous node attention layer, valuable information for subsequent prediction and analysis is extracted from the investment relationship network, syndicate relationship network, and cooperation relationship network, used to construct a prediction model, and obtain prediction results to assist the enterprise in more scientifically evaluating the enterprise operation status and future development prospects.

[0127] Optionally, the step of initially combining the investment relationship network information and the syndicate relationship network information through the hybrid layer module to obtain fusion network information includes:

[0128] Step A211, splicing the investment relationship network information and the syndicate relationship network information through a linear fusion formula and inputting them into a linear function layer to obtain a linear fusion matrix, and obtaining fusion network information according to the linear fusion matrix,

[0129] wherein, the linear fusion formula is:

[0130]

[0131] wherein, is the linear fusion matrix, and || represents the concatenation operation.

[0132] Specifically, taking the investment relationship network information and the syndicate relationship network information as inputs, when processing the investment relationship network information, methods such as the Graph Attention Network (GAT) can be used to extract node features and edge connection information from the investment relationship network. When processing the syndicate relationship network information, similar methods such as the Graph Attention Network (GAT) can also be adopted to extract node features and edge connection information from the syndicate relationship network, and then fuse the investment relationship network information and the syndicate relationship network information. Operations such as weighted sum, concatenation, and multiplication can be used to fuse the two network information to obtain the fused network information, which is used as the output for subsequent prediction and analysis. The purpose of the hybrid layer module is to fuse different types of network information to make full use of information from different sources and improve the performance and expressive ability of the prediction model. For example, in the capital market, the investment relationship network provides connection information among target enterprises, investment funds, and investors, while the syndicate relationship network provides cooperation relationships and industry experience among investment funds. By fusing the information of these two networks, the impact of investment relationships and syndicate relationships on target enterprises can be better captured. It should be noted that in the hybrid layer module, different fusion strategies and algorithms can be used according to specific requirements and problems. The fusion operation can be simple weighted sum and concatenation, or more complex operations, such as using an attention mechanism to assign weights to different networks, or capturing the interaction effects between different networks through element-wise multiplication. The fused network information can be further used for tasks such as prediction, classification, and clustering.

[0133] Optionally, inputting the fused network information and the cooperation relationship network information into the meta-path attention module, obtaining the meta-path weights and the node attention matrix, and fusing the fused network information and the cooperation relationship network information to obtain node embeddings includes:

[0134] Step A221, processing the fused network information and the cooperation relationship network information through the meta-path weight formula to obtain the meta-path weights,

[0135] wherein, the meta-path weight formula is:

[0136]

[0137] where is the meta-path weight, D is the weight matrix, q T is the semantic attention vector, and b is the bias;

[0138] Step A222, normalizing the meta-path weights through the node attention matrix formula to obtain the node attention matrix,

[0139] wherein, the node attention matrix formula is:

[0140]

[0141] Among them, is the influence weight, is the meta-path weight of the said one;

[0142] Step A223, through the node embedding formula, process the meta-path weight and the node attention matrix to obtain the node embedding,

[0143] Among them, the node embedding formula is:

[0144]

[0145] Among them, is the meta-path weight of the said one, is the influence weight, and M is the node embedding.

[0146] Specifically, the fused network information and the cooperative relationship network information are used as inputs. Through the meta-path attention module, the weights of different meta-paths can be learned and calculated. A meta-path refers to a sequence of nodes defined in a graph, which is used to describe the relationship type and path pattern between nodes. In a heterogeneous graph, multiple meta-paths can be defined to capture the relationships between different types of nodes. For each meta-path, through methods such as the attention mechanism, its weight in the network is calculated. These weights represent the importance of each meta-path for describing the relationships between nodes. Methods such as multi-layer perceptron (MLP) and attention mechanism can be used to calculate the weights of meta-paths. Using the meta-path weights and the input network information, the attention weights of each node are calculated. The node attention weights represent the importance of each node relative to the meta-path, and they can be adaptively learned and calculated according to the characteristics of different nodes and the weights of the meta-paths. The fused network information and the cooperative relationship network information are fused, and the node embeddings are obtained by weighting with the node attention matrix. Through weighted fusion, different network information can be better combined with node attention, so as to obtain a more comprehensive and accurate representation of node embeddings. The obtained node embeddings are used as outputs for subsequent prediction, clustering, visualization and other tasks. Through the learning and calculation process of the meta-path attention module, the important relationships and influences between nodes in the heterogeneous graph can be captured, and a more problem-oriented representation of the nodes can be obtained. The meta-path weights and the node attention matrix can help the network better utilize the information between different types of nodes and relationships, improving the performance and expressive ability of the model. It can be understood that in the meta-path attention module, different meta-path calculation methods and fusion strategies can be selected according to specific problems and requirements. Meta-paths can be defined manually, or automatic learning methods can be used to discover and calculate important meta-paths. In addition, other graph embedding algorithms and network structure learning methods can also be combined to further optimize the expressive ability of node embeddings.

[0147] Optionally, predicting each of the node embeddings to obtain a temporary prediction result includes:

[0148] Predicting each of the node embeddings through a prediction processing formula to obtain the temporary prediction result,

[0149] wherein, the prediction processing formula is:

[0150] U = LSTM(M t-k , M t-k+1 ,..., M t ),

[0151] wherein, U is the temporary prediction result, M t-k is the (t - k)-th node embedding, M t-k+1 is the (t - k + 1)-th node embedding, and M t is the k-th node embedding.

[0152] In one embodiment, this embodiment predicts the financing possibility of start-up enterprises. First, we can import relationship data from the CVsource and incopat websites to construct a relational network database for the model, and make predictions based on this data. When the target enterprise user (start-up enterprise user) opens the interface, they will be required to enter the company name, the financing time of the previous round, and the financing round. If the corresponding financing event can be queried in the database, the relevant events in the 36 months before the event occurred for this company will be scheduled, including historical financing events and patent cooperation events. The user will be asked whether the relationship information is correct and whether they need to modify the event occurrence time and the participating parties. If no modification is required, the prediction can be made directly. If modification is required, the relational network database will be updated after the modification and then the prediction will be made; if the corresponding financing event cannot be queried in the database, the user needs to manually enter the time when the previous round of financing occurred and who the participating venture capital VC is. According to this time, the relevant events within 36 months before this company will be scheduled, including historical financing events and patent cooperation events. The user will be asked whether the relationship information is correct and whether they need to modify the event occurrence time and the participating parties. If no modification is required, the prediction can be made directly. If modification is required, the relational network database will be updated after the modification and then the prediction will be made. After the prediction is completed, the interface will give the possibility of receiving the next round of financing and rank the relevant nodes that affect subsequent financing, so as to guide start-up enterprises to achieve effective relationship expansion. It can be seen that based on the CVsource and incopat data, as the information input of start-up enterprises increases, this database still has the possibility of update and iteration. As a VC user, after entering the company name, they can receive a monthly list of the latest high-quality start-up enterprises to assist them in making investment decisions. In addition, this model can also provide a list of closely related start-up enterprises to help them identify potential highly correlated start-up enterprises.

[0153] The advantages of a start-up enterprise risk assessment device described in this embodiment and a method for constructing an enterprise risk assessment model are the same as those of the prior art, and will not be elaborated here.

[0154] To solve the above problems, the present invention also provides an enterprise risk assessment method, including:

[0155] Obtain the heterogeneous network graph set of start-up enterprises;

[0156] Input the heterogeneous network graph set into the enterprise risk assessment model to obtain an enterprise risk assessment report.

[0157] The advantages of the enterprise risk assessment method and the enterprise risk assessment model construction method described in this embodiment are the same as those of the prior art, and will not be elaborated here.

[0158] In one embodiment, an enterprise risk assessment model construction device is provided, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above-mentioned enterprise risk assessment model construction method is implemented.

[0159] It should be noted that the device can be a computer device such as a server or a mobile terminal.

[0160] Figure 2 The internal structure diagram of a computer device, that is, the above-mentioned enterprise risk assessment model construction device, is shown in one embodiment. The computer device includes a processor, a memory, a network interface, an input device, and a display screen connected through a system bus. Among them, the memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and can also store a computer program. When the computer program is executed by the processor, the processor can implement the multi-capability collaborative operation method. The internal memory can also store a computer program. When the computer program is executed by the processor, the multi-capability collaborative operation method executed by the processor can be implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0161] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned enterprise risk assessment model construction method is implemented.

[0162] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided by the present invention can include non-volatile and / or volatile memories. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0163] It should be noted that in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0164] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features invented herein.

[0165] Although the present invention is disclosed as above, the scope of protection of the present invention is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will all fall within the scope of protection of the present invention.

Claims

1. A method for constructing an enterprise risk assessment model, characterized in that: include: Step S1, obtaining investment event data before a preset time and constructing an investment relationship network of the target enterprise; Step S2, constructing a syndicate relationship network of the venture capital according to the venture capital in the investment relationship network; Step S3, constructing a cooperative relationship network of the target enterprise according to the cooperative enterprises in the investment relationship network; Step S4, constructing a first investment network diagram, a first patent cooperation network diagram and a first syndicate network diagram of the venture capital of the target enterprise according to the investment relationship network, the syndicate relationship network and the cooperation relationship network within a first preset time interval, and generating a first heterogeneous network; Step S5, constructing a second investment network diagram, a second patent cooperation network diagram and a second syndicate network diagram of the venture capital of the target enterprise according to the investment relationship network, the syndicate relationship network and the cooperation relationship network within a second preset time interval, and generating a second heterogeneous network; Step S6, repeating steps S4 and S5, constructing a plurality of corresponding investment network diagrams, patent cooperation network diagrams and syndicate network diagrams according to the investment relationship network, the syndicate relationship network and the cooperation relationship network in different preset time intervals, and generating a heterogeneous network diagram set according to all the investment network diagrams, the patent cooperation network diagrams and the syndicate network diagrams; Step S7, training the preset prediction analysis model according to the heterogeneous network atlas to obtain an enterprise risk assessment model.

2. The enterprise risk assessment model construction method according to claim 1, characterized in that: The acquisition of investment event data before a preset time and construction of an investment relationship network of the target enterprise; Building a syndicate relationship network of the venture capital according to the venture capital in the investment relationship network; According to the cooperative enterprises in the investment relationship network, a cooperative relationship network of the target enterprise is constructed, including: According to the industry to which the target enterprise belongs, generating an industry characteristic matrix; According to the location of the target enterprise, a target enterprise location feature matrix is ​​generated; According to the location of the venture capital, a venture capital location characteristic matrix is ​​generated; According to the types of the venture capital, a type characteristic matrix is ​​generated; According to the location of the cooperative enterprise, a cooperative enterprise location feature matrix is ​​generated; Obtaining a complete node feature matrix according to the industry feature matrix, the target enterprise location feature matrix, the venture capital location feature matrix, the category feature matrix and the cooperative enterprise location feature matrix; Labeling the target enterprise, the venture capital and the cooperative enterprise to obtain a node label matrix; The investment relationship network, the syndicate relationship network and the partnership relationship network are constructed according to the complete node feature matrix and the node label matrix.

3. The enterprise risk assessment model construction method according to claim 1, characterized in that: The prediction analysis model includes a heterogeneous node attention layer, a heterogeneous meta-path attention layer and a sequence representation layer. The enterprise risk assessment model is obtained by training the preset prediction analysis model according to the heterogeneous network atlas, including: Step A1, inputting the heterogeneous network atlas into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information and partnership relationship network information; Step A2, inputting the investment relationship network information, the syndicate relationship network information and the cooperation relationship network information into the heterogeneous meta-path attention layer, wherein the heterogeneous meta-path attention layer includes a hybrid layer module and a meta-path attention module; Step A21, performing a primary merging of the investment relationship network information and the syndicate relationship network information through the hybrid layer module to obtain fused network information; Step A22, inputting the fused network information and the cooperative relationship network information into the meta-path attention module, obtaining the meta-path weight and the node attention matrix, and fusing the fused network information and the cooperative relationship network information to obtain node embedding; Step A3, obtaining all time series in the heterogeneous network atlas through the sequence representation layer, repeating steps A1-A2, obtaining the node embeddings corresponding to all the time series, predicting each node embedding to obtain a temporary prediction result, and tuning the preset prediction analysis model according to the prediction result and the heterogeneous network atlas to obtain the enterprise risk assessment model.

4. The enterprise risk assessment model construction method according to claim 3 is characterized in that: The heterogeneous network graph set is input into the heterogeneous node attention layer for independent learning to obtain investment relationship network information, syndicate relationship network information and cooperation relationship network information, including: Step A11, obtaining an investment relationship meta-path, a syndicate relationship meta-path and a partnership relationship meta-path according to the investment relationship network, the syndicate relationship network and the partnership relationship network; Step A12, obtaining the node weight of the first node to the second node in the investment relationship meta-path, the syndicate relationship meta-path and the cooperation relationship meta-path according to the node weight formula, Wherein, the node weight formula is: in, is the node weight, is the metapath Φ p The node-level attention vector under σ is the activation function, || represents the connection operation, and h i represents the feature vector of the i-th node, h j represents the feature vector of the jth node; Step A13, according to the standardization formula, the node weight is standardized to obtain the attention weight, Wherein, the standardization formula is: in, is the attention weight, is the node weight; Step A14, adding the influence weights of all the adjacent nodes of the first node, and repeating steps A12 and A13 multiple times to obtain multiple attention weights, splicing the multiple attention weights to obtain the node attention matrix of the first node under the investment relationship meta-path, the syndicate relationship meta-path and the partnership relationship meta-path, adding the influence weights of all the adjacent nodes of each node, repeating steps A12 to A14 multiple times, and combining to obtain a complete node attention matrix; Step A15, obtaining investment relationship network information, syndicate relationship network information and partnership relationship network information according to the complete node attention matrix.

5. The enterprise risk assessment model construction method according to claim 3, characterized in that: The initial merging of the investment relationship network information and the syndicate relationship network information by the hybrid layer module to obtain fused network information includes: Step A211, concatenate the investment relationship network information and the syndicate relationship network information through a linear fusion formula, and input them into a linear function layer to obtain a linear fusion matrix, and obtain fusion network information according to the linear fusion matrix. Wherein, the linear fusion formula is: in, is the linear fusion matrix, and || represents a connection operation.

6. The enterprise risk assessment model construction method according to claim 5, characterized in that: The step of inputting the fused network information and the cooperative relationship network information into the meta-path attention module, acquiring the meta-path weight and the node attention matrix, and fusing the fused network information and the cooperative relationship network information to obtain node embedding includes: Step A221, processing the fusion network information and the cooperative relationship network information by using a meta-path weight formula to obtain the meta-path weight. The meta-path weight formula is: in, is the meta-path weight, D is the weight matrix, q T is the semantic attention vector, b is the bias; Step A222, normalizing the meta-path weights using a node attention matrix formula to obtain the node attention matrix, Among them, the node attention matrix formula is: in, is the influence weight, is the meta-path weight; Step A223, processing the meta-path weight and the node attention matrix through a node embedding formula to obtain the node embedding, Wherein, the node embedding formula is: in, is the meta-path weight, is the influence weight, and M is the node embedding.

7. The enterprise risk assessment model construction method according to claim 3 is characterized in that: The step of predicting each node embedding to obtain a temporary prediction result includes: Predict each of the node embeddings using a prediction processing formula to obtain the temporary prediction result, Wherein, the prediction processing formula is: U=LSTM(M t-k ,M t-k+1 ,...,M t ), Among them, U is the temporary prediction result, M τ-k is the embedding of the tkth node, M t-k+1 is the embedding of the t-k+1th node, M t is the embedding of the kth node.

8. An enterprise risk assessment model construction device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the enterprise risk assessment model construction method as described in any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the enterprise risk assessment model construction method described in any one of claims 1 to 7 are implemented.

10. A method for enterprise risk assessment, characterized in that: include: Obtain heterogeneous network graphs of startups; The heterogeneous network atlas is input into the enterprise risk assessment model obtained by the enterprise risk assessment model construction method according to any one of claims 1 to 7 to obtain an enterprise risk assessment report.

Images